If you believe what you read on LinkedIn, real AI power uses are managing “teams of agents”. No, swarms of agents. What does this even mean? To cut through the hype, most of the time, all that means is having a few windows of Claude Code or OpenAI Codex doing tasks at once. Heck, if you’ve ever asked ChatGPT a question, left it to go research an answer, then done that 2 more times and waited for them both to finish, you’re arguably managing a team of AI agents. Jargon Debunked.

Whilst the people claiming to constantly manage 10 Claude Code agents at once are probably lying (notice how many of them are AI consultants or selling an AI course), there is actually truth to the hype. I personally often manage 3-4 agents at once, and know people whose default way of working is managing 2-5 Claude Code conversations at once.

But it’s not as easy as firing off a bunch of AI agents then sitting back and watching the productivity benefits roll in. Managing a team of agents is exhausting. Every agent you launch becomes a thread you have to mentally hold open until you’ve reviewed what it’s done. This post discusses a few ideas I have to reduce that mental burden, so you can do more and feel better.

The key problem we face here is context switching. Moving between agent chats requires constantly unplugging your mind from one task and plugging back into another....then swapping back again a few minutes later. This isn’t new. Knowledge workers have always had to do this: simultaneously answering their managers’ emails about a report whilst also writing said report and watching the football. And it’s always been mentally taxing. But managing a team of agents is plate spinning on a whole other level.

Tactics to limit the burden of context switching

Don’t be afraid to bin an agent’s output entirely

You can just delete conversations that you don’t have the bandwidth for. Even if it’s dozens of pages or hundreds of lines of code. It feels wasteful, but it’s more wasteful to burn precious human cognitive resources reading something that isn’t useful or valuable, or that you don’t have the bandwidth to use properly. You can always get more computing power, you can’t get more mental freshness (well, unless you count coffee and ADHD medication).

Batch decisions of the same type.

Don’t review one agent’s research report, then context-switch to drafting a social media post with another agent, then context-switch to evaluating a hiring shortlist. Group like with like.

Match the stage of agent work to your energy level.

Work with agents is often a four-step process:

1. You and the agent plan the task

2. They go do it

3. You review

4. You iterate together
   Step 1 requires the most cognitive energy, so do it when you’re cognitively sharpest. I’m a morning person, so that’s mornings for me, but you might be different. Save the simpler review and iteration stages for when you have less mental energy.

Warning! You can review when you’re tired, but don’t post anything. Don’t email the report draft to your manager, don’t click “post” on that important company LinkedIn update. AI outputs can look done when they actually aren’t, and a tired mind might think they’re ready when they contain embarrassing errors.

Name your threads

Rename your conversations to something useful, rather than relying on the name the agent gave it. A vague “Animal welfare research” becomes “Create research workflow for Sonia”.

Sometimes I open a convo and am like “ok I get what we’re doing here but why are we even doing this?”. The agent will default to saying what is in the convo, but often you also need why you’re having that convo in the first place to get your head back into it.

Pin your threads

Scrolling through a dozen threads to find the one that you were working on yesterday is a pain. I pin all threads until they are finished. That way I can simply skim the pinned threads, and either progress them, or if I’m done with them, unpin them.

Just ask the agent to keep going

The more that agents check in, the more context switching you have to do. Sometimes an agent will come back for feedback too early. It will stop working and ask you for direction, when really it could have kept going. Sometimes if I can’t be bothered to direct an agent I just ask something like “If it’s safe to do so, can you just make a reasonable assumption and carry on, or do you really need my guidance?” I’d probably say 1/3 of the time it will go back to solving the task.

Ask your agents to help you

When you jump back into a thread, don’t be afraid to say “catch me up on what we’ve talked about.” You don’t need to remember everything from scratch. The agent already has it. Use it.

You can also pre-emptively do this if you are working on a thread and know it’s going to be a while before you come back to it. When you finish the work, say “I am going to drop this task for a few days, write a summary for me when I return”

You can also ask agents to always give outputs in a format that’s easiest for you to process. It doesn’t have to be some ugly markdown document. Ask agents to make plans / status reports as nice HTML pages with interactivity and visual fun:

Perhaps the best place for you to review something isn’t even in your AI app. If you’re the kind of person who feels most comfortable reviewing things in email, perhaps because you have done so by managing humans for many years, then don’t be afraid to just ask AI to email its plan to you. You can have it turn the plan into a PDF or Google Doc and review it that way.
Tactics to limit the burden of context switching 3.png

Beware the temptation to “just fire something off”

It’s so tempting to just spin up another agent and throw another task at them. Sometimes you need to resist the urge to do this, because you just create an unmanageable mess of review work for later. It might feel very sci-fi to assign agents tasks before you go to bed and have them work while you sleep. In the morning, you’ve got to review it all! Ultimately you need to know how many things you can have in flight at once.

Accept that this work is genuinely taxing.

If I go deep on working with three or more Codex/Claude Code agents in parallel, I’m done for the day after four or five hours. I might do light work or chill out for a while, I often don’t work 8 hours those days. I’m exhausted! If you find yourself resting more during the work day, don’t beat yourself up about it. It isn’t a productivity failure, it’s the nature of the work.

What strategies do YOU use to manage more than one AI agent without becoming exhausted?

I want to give Claude a link to a Google doc and have it return a clean list of every comment, who wrote it, any replies, and the exact sentence in the document it refers to. Then Claude and I could work through the feedback together.

Basically, doing this is a pain that involves using Google Cloud, so is not worth the time for most people. Figuring this out was frustrating enough that I thought I’d write it down. That way, other people don’t waste the same time.

The native connector doesn’t work

Claude’s native Google Drive connector can pull the comments. It returns every comment, the author, the replies and the open/resolved status, all correct.

Separately, it can also pull the document text with the commented ranges marked up inside it.

What’s bullshit is that it can’t combine them! It can’t tell you which comment goes with which piece of text. The comments come back tagged with one kind of ID. The highlighted ranges in the document come back tagged with a completely different kind of ID. Nothing in the response links the two.

The dangerous thing is that when I first asked Claude to match them, it did: by guessing which comment went with which sentence. So obivously it got a bunch wrong, and then didn’t tell me in classic LLM fashion.

The raw API works but isn’t worth it

Connectors are just AI-friendly packages that contain API code. If a connector doesn’t work, the next best thing to try is using the API directly, having AI write all the code for you.

Google’s Drive API can give us what we want. It has a field on the comments resource called quotedFileContent, and its value is literally the text the comment refers to. So the correct way to do it is a single call. You run comments.list on the file, ask for quotedFileContent, content, author and replies in the fields, and every comment comes back already paired with the sentence it’s anchored to. No guessing. No parsing. No reconstruction (Manage comments and replies, Google; comments REST reference). Someone on Google’s own forum asked for this precise thing, the comments plus the highlighted paragraph, and that field is the answer (Google Docs community thread).

So the obvious fix is to skip the connector and talk to that API directly. Have Claude write a small Python script that calls comments.list with the right fields, then run it.

The problem is authentication. To use the API, the script has to use your Google account credentials. To make that possible, you go into the Google Cloud Console, create a project, switch on the Drive API, build an OAuth “desktop app” credential, and download a credentials.json file to sit next to the script. That’s just long enough for most people to say “fuck this”. But I’ve done this type of thing before. I hate Google’s Cloud Console but I’ll go there if I have to.

But it gets worse. if you go 7 days without using it....the token expires and you have to do it all over again (Google OAuth token expiration behaviour).

Google could probably fix this easily

Clever readers will be wonder: “If the Connector is just using the API, and the API has the answer we need, why can’t the connector call the right bit of the API?”

Because Google built the connector to not have access to the thing we need! The connector’s read call strips out the quotedFileContent field, which is the one field that would have solved this.

None of this is hard for Google. It’s one field. If Claude’s Google Drive connector either exposed quotedFileContent in its comment output, or preserved the link between comment IDs and the highlighted ranges it already returns, the whole problem disappears.

Summary

Basically, the connector can’t do it, the API can but you have to do a bunch of shit in Googles Cloud Concole and if you don’t use it every 7 days you have to go configure it again.

Can someone figure out a solution for this?

Claude Fable is the most powerful AI model on the planet by a large margin. You get some limited free usage of it for the next few days, then after that you have to pay for every conversation. People are scrambling for what to use it on.

I recommend 2 things:

Understand where Fable excels in your job so you know when to pay later

Find a hard, Opus level task, like a big piece of research, an app idea, a feature for an existing coding project. Then get Opus to do it on Max settings. Then have Fable do the exact same task on max settings. literally copy paste same prompt. Repeat until you're out of Fable.

This is useful because it allows you to directly assess how good Fable is for your hardest tasks. It allows you to figure out: Are there tasks where its a good idea to pay for Fable in the future? Or is Opus good enough?

Using Fable will be expensive. But for the hardest tasks, it might be worth it. But if you don't know how good Fable is for your work then you can't even begin to figure out whether it might be worth it.

Can't think of a task? Why not just look through older tasks you've already done? Arguably that's an even better comparison: retry real work with Fable and see if it's worth it.

Strategic thinking

Most AI’s these days are good at following plans. If you have a good goal, a great question, it will do the rest. It’s worse at coming up with those questions, identifying the right goals.

So use Fable to Refine your goals and questions. Not only will that give you strategic clarity, you can then just use less expensive models to answer your questions later.

Claude Cowork and Claude Chat have merged, use Cowork for almost everything now

The only thing 2 use cases for Claude Chat:

1. When you want to continue a conversation on the mobile app and don’t want to use Cowork Dispatch

2. If you’re on the Pro plan (the ~$20 one) and regularly run out of Claude. Claude Chat uses slightly fewer tokens to do the same task.

But other than than that, Cowork is better in every way.

Claude in Chrome is amazing….if you're on the Claude Max plan

So found out that if you’re on Claude Max, you get to use Claude Opus and Claude Sonnet in the Claude Chrome extention. But if you’re on free Claude or Claude pro (the ~$20 one), you just get the shitty Haiku model!

So unless you’re on Claude Max, I’d be wary of using Claude in Chrome for everything.

Offload the predictable parts of your skills and scheduled tasks to Python scripts: more reliable, fewer tokens

The core idea is to have Claude rewrite your skills or scheduled tasks to offload as much work to Python scripts as it can. Instead of having Claude do all the work, you have Claude run a Python script that does most of the work, and Claude just does the bits that require genuine flexible intelligence.

A brief explainer on how this works: when you run a skill or a scheduled task, what you’re doing is basically just sending a pre-written prompt to Claude. Then Claude goes to work; it might run a search, read a document, write some code and run it.

However, you can also have a skill or scheduled task run a pre-written Python script. Why would you want it to do this? Well, asking Claude to do a task is always going to be a little bit of a gamble. If you ask Claude to search your folder for information about a project, there’s always a chance it’s not going to find everything. If you ask Claude to search for information online 10 times, it won’t always search the same sites. This is a fundamental feature of Large Language Models; they are “probabilistic”.

However, a Python script is “deterministic”. It always does exactly the same thing, or it crashes. What people don’t realise is that there are almost certainly parts of your skill or scheduled task that actually can be done with a Python script. Instead of asking Claude to do the task, you have Claude run a Python script. The Python script does most of the tasks and then sends the information to Claude. Claude finishes the task. Using a Python script for some part of the task gives you increased reliability.

Example: imagine that every day I want Claude to search my inbox and collect all the emails about a specific project. Let’s say about plant-based pet food, and summarise those emails for me.

Every morning when this runs Claude is going to come up with a search strategy from scratch: it’s going to think of the right keywords, then it’s going to figure out how to use Gmail. Then it is going to figure out how I want the answer summarised.

But instead, we can give it a Python script that automatically pulls all the emails that feature the keywords “pet food”. Now Claude has a much easier and simpler job. Instead of figuring out a whole search strategy from scratch, it just has to summarise the emails that the Python script gives it.

A second advantage of doing this is that it reduces the amount of your Claude tokens. If Claude has to think of a strategy to solve a problem, then run three or four searches for a file, that uses up tokens. But instead, if Claude only has to run a Python script and then collect the results, all of the work has actually been done by the Python script, and Claude has done very little work. This means that Claude doesn’t have to spend as many tokens. This means that you can preserve more of your Claude usage for important tasks, and for tasks that genuinely require Claude’s intelligence and flexibility.

In 2026, LLMs got terrifyingly good at software engineering. So why are they so unimpressive at data science?

1. Data science doesn’t really use tests that much

LLMs are so useful for software engineering because of a powerful 1-2 punch:

1. As long as software tests are good, and code passes those tests, you can be somewhat confident that LLM code is good. You can have LLMs quickly produce vast amounts of code, and you don’t have to review it line by line.

2. LLMs are really good at writing lots of tests quickly.

On the other hand, data science success is fuzzier, and involves far fewer tests. Data science is mostly exploratory: you’re not building a product that either works or doesn’t work. You’re discovering what works best from a list of options. What is best is often not clear: whether a model makes a good prediction can depend which metric you prioritise (precision, recall etc). Some models might produce better results, but are harder to interpret. All statistical models make assumptions that you may or may not be willing to make. LLM struggle more with this fuzziness (as do humans!).

Data scientists usually don’t write tests unless they’re turning an analysis into a production model, and even then you could argue that its the job of MLOps (if you’re lucky enough to have MLOps people). I know many data scientists who never write tests and don’t need to.

2. Data analysis is iterative

Whilst LLMs are amazing at building software in Python, they are not as good at data analysis in Python. Analysing data is an iterative process. You’re exploring the data with tables, charts and statistical tests, and learning about the data. Then you use what you’ve learned to produce more informative tables, charts and statistical tests. Rinse and repeat until you either run out of time or feel like you understand what’s going on.

Today’s LLM agents can kind of do this, but it’s not that impressive. LLMs can code better than senior data scientists, but their data exploration is more on the level of a junior

LLMs have the components to be great at data science.

Bizarrely, most of the capabilities to do good data science are already inside the models. If you ask them what statistical test is best for a given situation, they give great answers. If you ask them to code up a specific analysis, they do a near-perfect job. If you ask them to critique your graph, they make helpful suggestions. But despite all this, they don’t consistently put all those skills together. Why?

I have 2 theories:

1. The training data for data science does not often contain objectively correct answers. LLMs are tuned using reinforcement learning, which basically amounts to “have it produce loads of answers then label those answers as either good or bad and have it learn the difference”. Code that doesn’t run is an easy “bad”. A data analysis that makes “arguably unrealistic assumptions” is harder to judge, so maybe the LLM gets a fuzzier signal. Data science blog posts and tutorials often follow a plan of “do whatever gets the best out of sample error”, which means that there often aren’t hard rules. Just ask your data science friends “Which of these models wins: XGBoost, Random Forest, neural networks or linear regression?”. They’ll tell you “it depends on a lot of things”. A model will struggle to distil clear principles for that kind of answer. Additionally, analyses by statisticians always contain a health dose of “we think this is probably the best analysis for this data but here are the many limitations”. Neither data scientists nor statisticians claim to have the answers to data problems, so LLMs probably struggle to identify the best practices.

2. Data science requires far more context than software engineering. Data analysis code can rune perfectly and be beautifully architected, yet produce an analysis that is totally wrong because it lacks context. For example, entries in the “price” column of a dataset generally should not be negative, right? …unless a business allows customers to return products and records them as “negative purchases”…or you’re describing the wholesale price of electricity from the national grid (but not if you’re a country with no wind/solar). And if the price of stock market futures is negative, its not wrong per se, but something quite dramatic is happening! And price should never be above £400 if you’re a fast fashion company, but always above that if you sell handmade Italian suits. The correctness of an analysis is entirely defined by the business context, so its hard for LLMs to identify the underlying patterns of a good data analysis.

Coding agents still struggle with notebooks

Data scientists make heavy use of notebooks (e.g. Jupyter) as an interactive way to analyse data. But while LLMs are exceptional at working with regular .py files, they are not nearly as good at using Notebooks in my experience. Claude Sonnet 4.5 can build a functioning interactive website in 1 prompt, but will randomly write notebook code cells in the wrong order! Even the breathtakingly good Opus 4.6 will do things like select the wrong Jupyter kernel to run a notebook.

This may be because Jupyter notebooks are long, convoluted JSON files under the hood. LLMs will probably get better at dealing with them, maybe even this year. But it may also be that the solution is to create a new file type that preserves human readability and interactivity with ease of use by an LLM. I have been trialling using hydrogen notation to have .py files that I can run as notebooks. Maybe we should all be looking at Marimo Notebooks? Or should we just be abandoning notebooks altogether and using a more LLM-native way to analyse data?

Where coding agents really accelerate data science

There is one key way in which LLMs are transformative for data science: iteration speed. I might have five hypotheses about a dataset, and some of those hypotheses could be tested in 2-3 different ways. I might generate 10 different charts for an in-depth analysis. I might have to deal with multiple types of messy data. That’s a lot of code to write! But LLMs are so good at writing code that they can fast-track all of that boredom. Using coding agents, I can test more hypotheses and iterate faster. The LLM is unlikely to give me the answer, but it allows me to arrive at the answer quicker.

A secondary way they help as a thought partner. Claude Opus 4.8 has the knowledge of PhD in statistics inside it, even if it doesn’t consistently apply that knowledge. I love chatting with Claude or Gemini about the merits and assumptions of different analyses.

How do you use LLMs to do data science? Have you overcome these limitations?

One of the problems LLMs have always had is that they are not fully reliable. Give them a list of 100 things to do, they will happily say they can do it. But then they’ll run out of attention after doing 60 of them, and won’t tell you. It might even claim it did finish the task! However, people have been telling me that this issue is not as bad as it used to be, so I re-investigated.

I had a pile of Instagram accounts I’d followed over the years and wanted to clear them out. 83 of them. I as dreading it, because Instagram makes it deliberately tedious. There’s no “unfollow everyone” button. You have to go one profile at a time: click unfollow, go back to your following list, find the next one, repeat.

So I just told Claude in Chrome to do it. And it did. It went profile by profile, clicked unfollow, went back to my following list, found the next person, unfollowed them. 83 times. Without stopping, without complaining. I expected it to get through maybe 20-30, then I’d tell it “carry on“ and repeat that a few times. But it bloody did the whole thing!

How I manage many Claude conversations at once: pinning and renaming chats

AI power users have multiple conversations going at once. How quickly and easily you can context switch between them determines how productive you can be.

Two ways I manage it:

1. Rename conversations If you think you’ll come back to them. I often rename them to the goal of the convo, like “Research and buy best vegan feta” or “fact check all references in report. That way when I jump back into that convo, I know what done looks like.

2. I like to ‘pin’ open conversations until they are done. Once I’m done with them, I unpin them. That way, when I jump into Claude, all my ‘open loops’ are at the top.

What do you use to help you manage multiple Claude conversations?

I saved a Claude in Chrome shortcut that auto-fills application forms for me

One of my favourite uses of Claude Chrome is filling forms for me. The only issue is that Claude in Chome is not attached to your regular Claude account: it has no memory of you! I think this is a security thing: if Claude Chrome had all the memories on your Claude account, a malicious website could trick Claude into sharing them.

To get around this, I created a Claude Chrome shortcut called /fill-form. These are just like skills: basically you type the skill name and it runs a custom prompt.

Now when I type it on a web page, it will run a prompt that basically says “You should attempt to fill out as much of this form as possible using the following information: Full name _____ DOB _____ address _____ ” I then provided my full name, date of birth, email address, phone number, regular address, and LinkedIn profile in the prompt. Claude will pick the info it needs and fill out the form!

I consider the security risks of this to be minimal; I don’t think outside websites that Claude reads can trigger it’s shortcuts. However, I’m not going to put my passport details in there, just in case.

But there’s a different emotional response I want us to talk about: fear.

It’s ok to be a bit scared of AI

When I first saw a Deep Research report way back in February 2024, my stomach felt like ice. As a researcher, I felt my job was disappearing in front of my eyes. Multiple times in 2025 the Bryant Research team genuinely entertained the possibility that AI tools would make us obsolete soon. In the end we leaned in hard, adopting the boldest AI policy in the animal protection movement, and now are using AI to produce vastly better research than before. I reckon Bryant is safe...for now? Who knows what will happen next year.

With the pace of AI development, some days I think “there has never been more possibility than there is now, so many new things are possible!”. Other days I think “Jesus Christ, can’t it slow down for just a week?”, or “What the hell is the world even going to look like in 2 years?”.

You can be excited and scared

I’m excited that anyone in the world with an internet connection can benefit from the power of GPT5.5. I’m scared that terrorists have it too.

I’m excited that scientists have incredible AI research tools. I’m scared that people spreading misinformation have them too.

I’m excited that AI can empower the animal protection movement to make real change for animals. I’m scared that factory farming companies have it too.

How should you feel about AI? I want to tell you that all your emotions are valid, and so are mine. If you feel positive and negative about AI at the same time, you’re not confused or misguided. This is the future we’re living in.

Claude can add tracked changes to Word documents.

I generally advise against using AI for serious writing. Low-stakes internal writing is fine, but for anything client/donor/policymaker-facing, you should be writing it yourself.

I believe the best way to use AI in writing is to have it propose changes to the text you’ve written, and then you can decide how to actually adapt those changes. Luckily Claude can do this, meeting you exactly where you are: in the doc! Just like a coworker.

For example, you could have it review a document adopting the persona of a European policymaker, a British Labour MP, or something like that, and leave comments about how this MP would judge how they would perceive the piece. Or you could have it go through and identify sentences that need a citation or further explanation.

If you use Google Docs, like many people, you have to download it as a word document. Unfortunately, Claude cannot add comments to a Google Doc (probably my only major frustration with Cowork).

Use Claude in Chrome to add things to your calendar

One of my favourite things that Claude in Chrome can do is add events to your calendar. Just open Claude in Chrome on the page that details your event, and say “add this to my calendar”. You don’t even need your calendar open, it will open it in another tab for you. Then go do something else (Lots of people tell me they don’t like Claude in Chrome because it’s slow - you’re not supposed to sit there and watch it!).

Note that on this page, it gives me an “add to calendar” option, but that downloads an ICS file, which is quite tedious to then import into your Google Calendar.

You might argue that this barely saves you any time. The purpose isn’t really to save you time; firstly, it removes a tedious thing that annoys me. Second, it adds detail to my calendar that I would never have taken the time to do anyway. Claude will copy the title, full descriptions, locations, everything. In the past I never bothered with that and just put in location, time and rough title.

You can even, as I did, ask in plain english for buffer time and it adapts. I do this a few times a week and it has never gotten an event wrong.

As a researcher in the animal advocacy space, Google Scholar is my bread and (vegan) butter. It’s how I find peer-reviewed articles that I can base my research on.

But there is tons of great work that is not conducted in universities or published in journals. This is often called “grey literature”. If you don’t know where to find this work, you might miss it! That’s why I’ve compiled a list of places that I go to when doing a research project.

If I’ve left off a good source, do let me know!

Organisations / websites that put out new research

• Faunalytics: Their original studies are all indexed by academic search engines, and their Research library mostly contains accessible summaries of published papers. But! It also covers reports, whitepapers and other scholarship, so worth a search.

• Rethink priorities: Sometimes, but not usually indexed by Google Scholar. Usually around the fundamental questions of the movement as well as blindspots (e.g. shrimp, insects, wild animal welfare)

• Food systems research fund: some, but not all projects end up published in academic journals. Very strong work on fixing the food system.

• Ambitious Impact (formerly Charity Entrepreneurship) publish their research reports onto the feasibility and impact of certain causes.

  • I have been told that they don’t publish everything, so if you’re wanting to be extra thorough (or are really desperate for references) you could email them.

• Animetrics do great work focussing on animal protection in neglected regions.

• Rooted Research Collective relatively new but promising. They haven’t put out anything publicly in the last year but are still giving public talks, so likely to see something soon!

• Tiny Beam Fellowship winners are sometimes missed by Google Scholar, and is great research on animal advo in low and middle income countries

• Bryant Research social science result consultancy exclusively serving the animal protection movement and al protein space

• The Good Growth Co similar to Bryant but work in Asia and do a lot of capacity building work

• Animal Ask do very rigorous research on the best tactics for animal advo orgs to pursue.

  • Currently not indexed by academic search engines, but I have told them about this and so this may be fixed at some point.

• Animal Think Tank do research on which messages are most effective at engaging the public on animal advocacy issues.

• The EA forum contains generally high quality writing.

• Mintel is a leading market research company. Most reports cost $ but there are some for free here

• YouGov is a leading polling company. Like Mintel they often charge for their reports but you can often find polling for free here. You can also find more detailed reports if YouGov did them on behalf of a client that chose to make the results free, as Compassion in World Farming did here. I have no systematic way to find these reports, if you do, let me know! Sometimes googling “YouGov” AND “what you’re interested in” works.

• Our World In Data publishes accessible data and analysis on animal welfare.

• Animal Charity Evaluators maintains a guide on the evidence for and against various campaign types in their “Better for Animals” guide here. I have created a NotebookLM chatbot that allows you to chat with this fantastic resource here.

Organisations / websites that have done research but have stopped posting new research

Here you’ll find research that may not be indexed by Google Scholar. These places have either stopped doing research, haven’t posted in over a year, or shut down, but their research is still valuable

• North Mountain Consulting are an alt-protein consultancy that appears to be shut down.

• Open Philanthropy (Now Coefficient Giving) was posting farmed animal research up until 2021

Curated databases of papers

The research here tends to be less novel, and you can usually find this stuff on academic search engines. But they can be useful for zoning in on literature quicker.

• Inside Animal Ag are a great research summarizer

• Faunalytics Research library

• Animal data project

• ACE’s Research Library

Perhaps about once a month someone messages me on LinkedIn or the Hive Slack community, asking how a data scientist/analyst/engineer can use their career to help animals. This is something I’ve spent nearly 5 years thinking about! I originally trained as a data science after my PhD because I wanted to build skills to make the world a better place. These are my reflections on what I’ve learned, and conversations I’ve had with others.

Definitions

• Data professional: umbrella term for data scientists, data analysts, data engineers, statisticians, MLOps people, analytics engineers blah blah….

• I’m going to use the term “animal protection” to describe both animal protection non-profits and alternative protein companies. Most alternative protein companies don’t consider themselves to be part of the animal movement (or wouldn’t admit it publicly), but my assumption is that the reason you the reader would consider going into alternative proteins is because you want to help animals. So I’m just going to lump them all in together.

Paradoxically, while there is demand for data science in the animal protection movement, there is little demand for data professionals in animal protection organisations.

The bad news

Simply put, most orgs cannot afford a data professional. Not only that, they don’t have enough data, analytics or machine learning (ML) use cases to need even part-time data person. This includes both non-profits as well as alternative protein companies (who are usually small startups). I typically see 1-3 data roles a year advertised across the whole movement. There are probably less than 15 data scientists in paid employment in the entire movement.

How can organisations need data science but not data scientists?

Across all organisations, animal or no, for profit or non-profit, there is growing need for data skills. However, this doesn’t necessarily mean there is a growing need for dedicated experts like data scientists. Generally, it’s a growing need for everyone to have data fluency. Here are some ideas off the top of my head about how “non data people” might leverage data, machine learning and AI:

Data science skills and use cases for animal protection organisations

More generally:

• Any role that has to use spreadsheets at some point would benefit from data skills. If you have spreadsheets, you have data. Some data skills make you better at using those spreadsheets, and if you know R or python you can take that data out of spreadsheets and do much more with it (or more likely, direct AI to do it).

• Literally everyone needs to learn to use generative AI, because it is already impacting basically all jobs.

• AI models can do data analysis, but often make questionable analytical decisions that only a data person would catch.

More specifically:

• Marketers might upskill in AB testing

• Researchers or Analysts might learn to scrape websites.

• Researchers would also likely benefit from statistical training (which data professionals usually have) to do better “guesstimating”: estimating the size of a market, or the potential impact a program could have.

• If you monitor your organisation’s socials or comms, then social media analytics is important to understand what’s doing well and what isn’t.

• Product/project managers, as well as executives at medium to larger organisations would often benefit from dashboards and automated reports, but might only need a few.

• Fundraisers might pull together enough data on donors to do some cluster analysis to understand what types of people donate to the organisation.

• Website/app developers might add simple AI features into apps and websites

• Any organisation that wants to get serious about evaluating the impact of its programs and campaigns would benefit from “causal inference”, a discipline from economics that is increasingly being brought into data science.

I’m a data professional that wants to help animals. What are my options?

There are broadly 3 key ways a data professional can directly participate in the animal protection movement and one indirect way:

1. Doing a non-data job where you can use data skills

2. Fractional roles

3. Consulting

4. Earn to Give (spoiler: this is the main one I recommend)

Use your data skills inside a non-data job

As outlined above, plenty of roles within animal protection/alt proteins would be enhanced with some data skills. You could be one of those people! Whereas outside the movement you might be a data analyst working in the marketing department, in the movement, be a marketer that can do data analytics.

It may feel strange or even sad to train as a data scientist, only to not “be” a data scientist. It may even make you feel like the training was pointless. But I would advise placing your identity as an animal advocate first, and focus on your commitment to helping animals. Ask yourself not “how can I help animals as a data professional?” but “How can I help animals with my data skills?”. Or perhaps you could create a new identity of yourself as a “hybrid data scientist/marketer” or “data-driven fundraiser”.

Fractional roles

Some people work part-time at one organisation and part-time at a different one, splitting their work week between 2 organisations. For example, in early 2024 I worked 3 days a week at Peak AI as a data science consultant (non-animal) and 2 days a week at Bryant Research as a data driven researcher (an animal protection research org). Then, I split my week 50:50 between Bryant and being the Director of Data and AI at Greener by Default (I’ve since handed off the role but remain a part time data consultant for them).

Fractional roles split it down even further: You might work at 3,4 even 5 organisations. Each one only needs you (and can probably only afford you) for 1 day a week, or 1 week a month, or any other arrangement. This is likely to be more possible in the animal protection space than many other areas of the economy because organisations are all allies, rather than direct competitors. You could never work for Adidas and Nike at the same time, because they’re direct competitors. But you can simultaneously work for two organisations that both want to end factory farming. It would even have the indirect benefit of increasing potential for collaborations. At Bryant Research, some of our Bryads also work in academic labs or campaigning orgs, and we all benefit from being closer to each other (they’re not data scientists but the logic is the same).

It seems to be common in the animal protection movement to have multiple part time roles, so orgs are likely to be open to a fractional role arrangement if it benefits them. If you have your eye on a few organisations, I advise speaking with a few of them about this option and making them aware of it if they are not.

Consulting

Another option is for the animal protection movement to have consultants and consulting firms that concentrate expertise and then do contract work for various organisations.

Currently, when it comes to specifically data, there are a few consultants, but no consulting firms. I think the movement could always use strong data consultants, but don’t think there’s enough demand for a consulting firm specifically for data. However, a consultancy network where data consultants share work and help each other might be useful (I give this a mild but not trivial chance of success).

I’d probably not call yourself a data consultant (or similar) though. I would aim to become a Monitoring and Evaluation (M&E, sometimes called MEL) consultant or a research consultant that specialises in data-heavy projects. The main analysis needs of orgs are data-heavy research projects, and using data to measure the impact of their campaign for funders. The hard truth for data people here is that there are already research consultants and MEL consultants that have domain knowledge you won’t have, and probably enough data skills to do a reasonable job. You’ll have to put the time in to build your skills in MEL or research.

I won’t repeat the standard generic pros and cons of being a consultant. The disadvantage compared to fractional roles is that consultants are external to the organisation and so do not know all the organisational context and culture. This can make them slightly less effective at delivering projects if those projects are not well planned and the consultant is not good at discovery and scoping (i.e. figuring out the real problem to solve, not naively solving the problem the client puts in front of them). You also don’t always feel like part of a team which can be a bit lonely.

It is worth noting that while consultants typically earn good money, this is not the case in animal protection. The animal protection movement is terribly underfunded. Expect to earn significantly less than if you were a consultant for for-profit companies (as in, like 20-40% less!). That said, if you want to earn lots of money, you probably wouldn’t be in this space in the first place.

Earning to give

This is an indirect way to help animals, but honestly, it’s the main one that I would recommend if you’re a data professional wanting to help animals. If you work in data, you can earn very good money in the for-profit sector.

So I advise that you do that, and then donate 10-30% of your income to animal orgs. If you are a high-paid data engineer or machine learning engineer, your donations might be enough to pay someone in the movement’s entire annual salary. The animal movement is one of the least well-funded social justice movements, so donations are often needed far more than skills.

This is probably not be a satisfying answer for many people: donating doesn’t feel the same as contributing directly. But similar to what I said above, always keep in mind that the animals must come before our own egos, and if this is the best way you can have an impact for animals, then you should feel very good about yourself for doing it. I personally have huge respect for people who donate significantly to the movement.

Find more about Earning to Give here.

What about AI consulting and AI roles?

I intentionally left this out of earlier versions of this post but people keep asking me about it, so I’ll address it here. Some people feel like a data scientist joining the movement as a generative AI consultant is a natural fit. I don’t think it is.

I definitely think the animal movement needs more people with AI skills, and if you’re a data person thinking about retraining as a generative AI person, I think that’s a great idea.

However, I want to emphasise that I strongly believe that data people have little to no advantage when it comes to retraining in generative AI. AI before ChatGPT and AI after ChatGPT are technically similar tech, but in practice they are fundamentally different. If you’ve built AI solutions for a decade, I don’t think you’re any better placed to be a generative AI consultant than someone who’s been a marketer for 10 years or someone who’s worked in HR for 10 years. In fact, I personally think the people best placed to move into generative AI consulting are operations people and management consultants (I plan to write about this).

So while I would recommend people take this path, I wouldn’t say this is a particularly promising path specifically for data people. Be aware you’ll be starting from the same point as everyone else.

Conclusion

Animal protection could benefit a lot from data science skills, but lacks the funding and scale to make hiring dedicated data professional feasible. If data professionals want to make an impact for animals whilst also making a living, they might consider non-data science roles that use data science skills, fractional roles if they can get them, or staying in the for profit sector and donating.

Acknowledgements

Thanks to Antonieta Pais Nunes Lopes for inspiring this post. Also thanks to Kyle Behrend and Sam Tucker for valuable insights and comments.

Will AI automate all of our jobs? Or will it create tons of new ones? This is one of the biggest debates of the last few years, and I suspect we’ll be talking about it for years to come. There are good arguments on both sides and to me, it’s unclear what’s going to happen.

But one area of the job market is already showing signs of trouble: Entry level jobs. A recent report from Ravio shows hiring rates for entry level jobs in EU tech companies are down 73.4% in 2025, compared to just 7.4% in the rest of the job market. Employment prospects for graduates in the US have “deteriorated noticeably”, writes The New York Times. And Dario Amodei, CEO of Anthropic (who make Claude), recently warned that the technology could soon eliminate half of all entry-level office jobs.

What’s causing this?

Simple: when managers and senior employees learn to use AI effectively, they have less need for junior staff. Junior jobs usually consist of simple, repetitive tasks that don’t require a huge amount of expertise. AI is great at those.

Junior researchers spend hours crawling through large datasets to find the right stats for a report. They read dozens of Google search results and pull together “briefs” for senior researchers. They create powerpoint slide decks. But in 2026, you can do all that yourself with Perplexity, ChatGPT Deep Research and Gamma.

What I’m seeing on the ground

I’ve already seen it at both the organisations I work for:

1. At Greener By Default, until recently, I was the only member of the data team. In 2025 I managed to serve nearly all of our clients and build a data science codebase from scratch by making full use of the latest AI coding tools. In fact, I estimate in 2025, AI coding assistants increased my output 3-4 times! We were going to hire an assistant for me, but decided it wasn’t necessary. Instead we hired a Data Director, a more senior role focussed on strategy.

2. At Bryant Research, we get lots of people wanting to intern/volunteer with us. But I am torn: we usually can’t think of projects to put them on! For most projects, the additional admin and supervision an intern requires is not worth it; our researchers produce better research with AI assistance instead.

What do we need juniors for?

I wish I could write the rest of this post on “here’s why replacing junior employees with AI will be terrible for your organisation”. But I am struggling to do that. Consider the main weaknesses of AI:

• AI does not have your organisational context like an employee does.

• AI is not totally reliable.

• AI cannot take initiative and needs to be told what to do

• AI does not bring any “career capital” to a job (e.g. it does not have contacts or past life experience that can be tapped)

• AI doesn’t have a body so it cannot take physical action in the world.

Apart from that last one, all of those drawbacks apply to junior employees as well! I’d argue more.

In addition, AI models like ChatGPT have key advantages that juniors do not have. Namely that they have PhD level knowledge in most things humans have ever written about on the internet, and can produce that knowledge in seconds.

This sounds harsh, but in the age of AI, the day-to-day value of juniors/interns is not as certain as it used to be.

A poison apple for the animal protection movement?

For cash-strapped non-profits and scrappy grassroots campaigners, the temptation to replace junior roles with AI can be just as strong as in profit maximising corporations. If AI allows us to do more good with fewer resources, how can we say no?

And whatever you believe personally, what will your funders say? Some would argue that it’s unethical to use donor dollars to pay junior activists, when you could just use AI and divert those dollars to impact elsewhere.

This is also a problem for grassroots organisations who have very minimal paid staff. Senior organisers are increasingly facing a trade-off: Do you delegate tasks to a volunteer, who can only give a few hours a week around their life, or to AI who will hop to it the second you open your laptop? Sure, AI can’t go door to door knocking or attend a protest, but it can research campaign targets, produce social media content, identify news outlets for work and draft emails to companies. It can do most of these things as good as paid activists, let alone juniors.

But if the temptation is just as strong, the risks of this strategy are far worse for the animal movement than the for-profit sector. If Facebook doesn’t contribute to training new software engineering talent, it doesn’t matter: they will always have plenty of top quality mid and senior candidates to choose from whenever they want to hire. But the animal advocacy movement doesn’t train junior activists, where will the skilled, mid-level activists of the future come from? Who will fill mid-level openings?

The solution

To be honest, I’m not really sure what the solution is. I just see a big problem and want to start the conversation. I think leaders in the animal protection movement need to be talking about this.

• Should funders be earmarking money for apprenticeships?

• Should larger, better resourced organisations bear the “burden” of training?

• Should we pool resources and develop movement-wide training programs?

• Should we as a movement simply accept this as the “new normal” and focus recruitment on professionals that already have the skills we need?

• Can we find roles for juniors/volunteers where they can provide genuine value over AI?

• Could AI itself help provide the training / mentorship to upskill the next generation?

Last year, Bryant took on a fellow from The Reducetarian Foundation, Omar. The “deal” with this fellowship is they send a talented, pre-vetted college student trained by them to work for you. In return, you mentor them and give them real world experience in the movement, and help fund their salaries.

Don’t get me wrong, Omar was genuinely brilliant for his age and experience level; great analytical mind and critical thinking (he was also fun just to chat with!). When I weigh the hours I spent training him against what he produced for us, it seems like a decent “return on investment”.

But could I have got better quality work in less time, using AI on my own? Probably.

So why did we do it? At Bryant we talked about it internally, and decided that it’s important to us to support the development of the next generation of research for our movement, even if the ROI might dip negative. That sounds heroic, but I am still undecided on whether that was the best call. Time spent mentoring a junior could have been spent producing better work for our clients, allowing them to help more animals. Or we could have passed those cost savings onto our clients, giving them more resources for vital campaign work. I have also had conversations with other leaders whose intelligence I respect, who told me they don’t think this is our responsibility, and that we should focus on doing research.

With a bit more training, experience and encouragement, I could see Omar becoming a fantastic researcher, worthy of a job at Bryant or any other research org in the animal advocacy movement. But will the movement give him the chance? Can we afford to?

Further reading

This is really well explored in book-length in The Skill Code by Matt Beane

Thanks to Tobias Leenaert (ProVeg International) and Lauren Mee (Animal Advocacy Careers) for feedback on this post, and to Karol Orzechowski and Meghann Cant (Faunalytics) for interesting discussions on this topic. None of these people necessarily endorse any opinions in this post.

There are plenty of websites where people have shared their Claude skills. Don’t just copy them! You should always build your own skills. One of the best things about Claude’s skills is that they can be made ridiculously tailored to you and your workflows. So, you should try to make them as personalised as possible. Overly generic skills are less useful to you.

If you see a skill online that you think is useful, what you should do is read the skill.md file to understand how the skill works. Then, simply describe your own version of that skill to Claude and have it make the skill for you.

for example, let’s say you find a skill that opens your LinkedIn and triages new connection requests.

Claude (and all LLMs) still suck at graphs

In AI there is this concept called “The jagged frontier”, which refers to the fact that AI is incredible at somethings and crap at others.

A great illustration of this is in software engineering and data science. Both require coding in python, but whilst Claude is better at most software engineering tasks than anyone you or I know, it’s just not that good at data science. Because it doesn’t “see” graphs it makes (unless you directly tell it to) it tends to spit out shitty graphs.

I was recently analysing my preference in coffees from a dataset of 68 tastings I’ve built up over time (don’t ask). It made this:

Don’t blame Claude for the pretentious art deco colour palette. But why does it have vertical gridlines? and WHY did it put that in the title? What purpose do they serve? Why has it put in a faint “mean = 3.16” ?

It also made a stacked chart. Whilst I do see these online, they are generally considered a bad idea because its actually really hard to compare any segments except the bottom one. Sure, we can easily compare the blue bits, but comparing the green segments is quite hard.

I’d expect better graphs from a junior data scientist, let alone a model that can write python 10x better than me and 100x faster.

Claude in Chrome catches broken links and typos on your website, and you can schedule it weekly

Our Comms person at Bryant, Elise, has been using Claude in Chrome to comb through the Bryant website looking for broken links and typos, and it seems to be working pretty well! Try it on your own website.

Pro-tip: set up a Cowork Scheduled task to repeat this task weekly: it’s like having an intern check your website for errors every week. The way to do this is:

• Have Claude do it once

• Iterate until it works well.

• At the end of that conversation, ask Claude to make a repeatable skill (something like “scan-website-for-errors”) distilling the best practices, pitfalls and such from your whole conversation. Save that skill

• Ask Claude to set up a scheduled task that runs every week at whatever time that basically runs the skill and then tells you what happened. You could have Claude slack you, email you or just alert you in the Claude app. Whatever works for you.

This month I took a lot of built up annual leave: I did a 6 day dance festival in Albania with my wife, took my dear mum to see Pompeii for her 60th, attended the AVA conference in Toronto and EA Global in London. I had Cowork manage my entire month in one project, and it worked amazingly well! Things I did:

• Claude helped with the planning, such as advising where to stay in places to be close to events, what time to get flights to minimise jet lag, what to pack etc

• I gave it all my flight and accommodation details

• Claude acted as a tour guide to Pompeii, Herculeneum and Toronto, suggesting attractions, providing historical facts and making schedules

I structured it by making a folder on my computer and putting everything in there: pdf downloads of flight and accom details, conf tickets. I also had Cowork create a bunch of documents: itineraries, packing lists and such. Cowork was uniquely good for this because it can look at everything in a folder and create/update new docs on the fly.

Usually I didn’t have access to my laptop, so I kept it on at home (and plugged into the charger!) and used Cowork Dispatch to message it from my phone.

I’m aware this could have gone totally wrong, but I wanted to push Claude to it’s limit. Shockingly, Claude did not make a single mistake that I can detect!

You can manage multiple conversations with Dispatch

AI power users often have multiple convos going at once. That was not possible with Claude Dispatch before: you could only have 1 convo at a time and it was tedious.
That’s fixed now! Dispatch can have and maintain multiple convos

Claude often tells me that Claude Chrome is not available

This is annoying: half the time I tell Claude “use Chrome to do X” it says “the chrome tool is not responding”. I can’t quite figure out why, does anyone also get this? Has anyone solved this?

Crucially: Claude only gives you a certain number of tokens per 5 hours. There’s also a weekly limit. You have probably had the frustrating experience of “running out of Claude”. That’s because you’ve spend your token budget. Because some things lots of tokens and some things take few tokens, you might find that some tasks drain usage quicker than others.

More cool things Claude Chrome can do

The simple one: Summarise clickbait substack posts and tell you if they’re worth reading! Like “What unique analysis does this post do on the governments new farming policy?” .

Note: I won’t be offended if you have Claude Chrome summarises all my posts!

Once you archive a Claude co-work project, it’s very hard to get it back

The title really says it all, to be honest.

If you right-click on a Claude Cowork project, you can archive it. That does not delete the project but it stores it deep in your computer, making it hard to retrieve. The Claude app doesn’t let you unarchive co-work projects!

If you’re feeling a bit brave: you can use Claude code to get them back: you have to locate the a JSON settings file in your computer that decides what shows up in the Cowork UI.

Claude skills can contain python code…but will break in Claude Chat

claude skills are one of the most powerful features of Claude Full-Stock. The Bryant team has skills for everything. One advanced use of skills is that you can have a skill run a Python script. For example, you could make a skill where if you give it a dataset, it will read that data into Python, perform some analysis, and then output the results. It can also run some Python code to create a graph and display the graph. This capability is really powerful. However, one thing to watch out for is that if your skill contains a Python script, ClawedCowork can run it but ClawedChain won’t be able to execute it. Chat cannot handle it; it will just silently break.

Turn off connectors by default to save usage

If you’re not on the expensive £100-200 plan you probably run out of Claude often. If you’ve connected Claude to a bunch of other apps (and I recommend it) then what you might not realise is that they have a cost! Claude loads them at the start of every chat which can eat into your token allowance (see explainer above). If you find yourself running out of Claude a lot, it’s worth turning these off by default and turning them on when you think a conversation will need them.

Advice tailored to privacy-concerned professionals

Whilst there are brilliant new-generation AI presentation makers like Gamma, not everyone is allowed to use any tool they find on the internet. Many organisations have an “approved AI tools” list for security reasons, and getting a new tool onto this list can be as tedious as making the presentations in the first place!

So I thought I would come up with a way to significantly speed up making PowerPoint presentations using only tools that are probably already approved by IT: ChatGPT, Gemini and Claude. Pick one. I’ll refer to them collectively as “AI”

Tip 1: Turn any content into a slideshow outline

Quite often, we are giving presentations of other content that we’ve already created. For example, we might be doing a webinar on a report we wrote, or giving a conference presentation based on our academic paper, or pulling together several blog posts into a presentation.

In all of these cases, we’re converting long-form content into a concise set of bullet points that we can put onto PowerPoint slides. This process can be time-consuming and cumbersome, so let’s just ask AI to do it for us:

Tip 2: AI can literally create a PowerPoint for you now

Many people don’t realise this, but ChatGPT, Gemini and Claude can all create decks. Claude and ChatGPT make PowerPoints (you can also install a Claude extension for Powerpoint), and Gemini obviously makes Google Slides.

So once you’ve done a piece of work, had AI turn it into a presentation outline (see tip 1), you can whip it up into a deck in minutes.

I have created a Claude Skill that not only makes a deck following my branding, but it also uses an image generator to place images on every slide! This is a bit more advanced for this article, but I’ll probably blog about it in future.

Tip 3: Make bespoke, on-brand images

You don’t need a something as fancy as a custom Claude skill though. You can just do it “the old fashioned way”. Open up Gemini or ChatGPT (Claude has no image generator) and ask it to make you a few images. You can provide the text of a slide, briefly explaining what you want to communicate and ask for picture ideas. Then, you select one and have the tool create it for you. Gemini has an image generator built right into Google Slides if you’re on a paid plan.

However, you can go one step further: If you define your branding and tone, AI will make all the images in that style.

If you look though this slide deck I recently presented, you’ll notice they follow a consistent tone. My personal brand is inspired by art deco; gold, navy, forest green. A bit geometric, stylised people. Always generating images in this style makes my presentations coherent and uniquely me.

How do you do this? Before asking ChatGPT or gemini to generate an image, talk with it about art. Discuss art styles, give it examples of what you like, and then ask it to come up with a prompt to capture that style

Tip 4: Use AI to critique slides

Your slides are more than just content and pictures; you need to tie them together in an engaging way.

To do this: screenshot an entire slide, upload it to AI as a picture, and then ask it to critique it:

Ask it for the same feedback you’d ask a colleague for, like:

• “Is this slide too cluttered?”

• “Should I put the picture at the top or the bottom?”

• “Does yellow text on a blue grab attention or is it jarring?”

I can’t emphasise this enough: any time you’d ask a colleague for feedback, ask AI first. You’ll be surprised at how often it can provide helpful advice.

Conclusion

Here I’ve shown how you can Use AI to dramatically cut the time taken to make engaging presentations using only ChatGPT, Claude or Gemini. Try one of these tips today and cut your next presentation time in half. Then spend that time doing something more valuable, or at least more fun.

This could be automations in Make.com or n8n, or it could be “skills” in Claude Code. It might involve jargon MCPs, or the sinister-sounding OpenClaw. It could be Python scripts or cron jobs.

Btw, if that last paragraph made zero sense to you, don’t worry! This post is still for you. No jargon from now on.

In the last six months, something I’m learning is that often the best thing to do is NOT to solve the problem. Sometimes you should just wait. The models are developing so quickly, with new features added literally every few weeks. Why go through all the effort to create a bespoke system to solve a problem when you could just wait for Google or Anthropic or some random open source developer to solve the problem for you?

For example, people love the idea of paying automation consultants for bespoke workflows to handle their email inbox: triaging important emails, reminding you at the right time, even drafting responses. But I have so far held off on building one of these for myself or for any of my colleagues. I am pretty sure that Google is already trying to build most of this functionality into Gmail. I would bet that it will be ready this year. It’s not worth me spending days tinkering with a custom solution. Google will solve it for me.

When not to build

• When a great feature in an AI product is buggy AI companies are releasing new features all the time but they usually aren’t polished in the first few weeks. Anthropic released Claude Cowork in February, And while it has some bugs still, issues get solved every week. Avoid spending human hours trying to work around these bugs.

• When the project is too difficult. There are so many easy, powerful AI use cases out there, that I am extremely wary of committing to any AI project over 4 months. In 4 months, AI might have progressed so much that it could finish my project in a week. I have literally seen this happen several times.

• Projects that only try to reduce the cost of current AI workflows. I get it. AI is expensive and we’re not all made of money. Trust me, I work with NGOs for a living. That said I think that the vast majority of people are underspending on AI. They’re focusing on the costs and not the benefits. Every AI project has a return on investment: the benefits you get out, divided by the cost you have to put in. You should be focusing on getting more out not reducing the cost in. Plus the cost of these AI models is actually going down over time. Why put effort into reducing costs when you could just wait for the cost to come down naturally over time?

When to build

That said, here’s when I think a custom solution is best:

• When the benefits are so huge the ROI is positive even after 2-3 months. Last year I built a custom python automation to create slide decks for my presentations. Now Gemini, Gamma and Claude Cowork all do a great job and I don’t need it. However, the 2 days it took me to build saved me many more days of awkwardly moving text boxes around.

• When you need a highly customised solution that is unlikely to be provided by the market. It’s a safe bet that Google is trying to bring more AI features into Gmail. On the other hand, I don’t really know if it is a priority for Google at all to integrate Google Scholar into Gemini. If I need that (and I really do!), it might be best to build it myself rather than hoping.

• When you need a first mover advantage. lots of software development companies built their own custom tooling around AI coding. I’m willing to bet that most of the companies no longer use the tools that they built, they just use Claude Code or Cursor instead. However given how highly competitive the software industry is, it may have been worth it to not wait for the market to solve the problem. By building your own system today you can get the jump on your competitors tomorrow.

Stop building tools that will be obsolete by next Tuesday. Master strategic laziness instead.

• Claude sub-agents prevent overload: Spawn fresh sub-agents per subtask to keep context clean. Claude often skips this unless told explicitly.

• Claude’s memory is a killer advantage: Say “add to memory” mid-chat and it surfaces fluidly later. Gemini and ChatGPT can’t match it.

• Claude nonprofit team plan: ~$45 more for ~6.25x the usage. Lets you run Opus on everything.

• Gemini memory workaround: Build a Gem that reads a “Things I’m working on now” Google Doc. Edit the Doc to update memory.

• Gemini @drive connector: Separate write-up.

• Make.com blueprints as JSON: Dump into an LLM to debug or translate to n8n. Save often, Make.com can wipe unsaved work.

• Make.com debugging via Claude Chrome extension: Point it at a broken canvas and let it diagnose from screenshots when the JSON isn’t enough.

• AI connector security: Prompt injection risk only matters for connectors exposed to untrusted inbound messages (email, public Slack, LinkedIn). Internal-only connectors (Drive, Asana) are much safer.

1. Claude

Sub-agents prevent Claude getting overloaded

Sub-agents are one of the more underused patterns in Claude. The idea is that instead of having a single Claude chat do every step of a complex task in one long conversation, Claude spawns fresh copies of itself to handle individual subtasks. Each sub-agent gets its own clean context window (memory), goes off and does its thing, and only returns the final results to the parent. The parent never has to deal with all the intermediate noise.

Kevin’s morning-brief skill is a good illustration of the problem they solve. The skill is supposed to check Slack, pull out action items from Gmail, look at Asana, draft replies, and produce a morning briefing, all in one go. The more it was engineered, the worse it performed. The issue is probably context overload: by the time Claude is looking at Asana, its context window is already full of Slack and Gmail noise, so its attention to the Asana task is degraded. Ideally each of those data sources would be read with a fresh context window, and only the cleaned-up result would be passed back to whichever agent is writing the brief.

Hilariously Claude will sometimes say “I will fetch this in parallel”… and then not use sub agents. Basically it will lie to you! You can verify after the fact by asking Claude directly whether it used sub-agents. Claude does not always use sub agents by default and often avoids it for some reason. You need to explicitly tell Claude to use sub agents.

Side note: Claude Code and OpenAI Codex (it’s coding model) both also have sub agents. Codex gives each sub agent its own name, naming them after famous scientists. Cute.

Claude’s memory is it’s killer advantage over Gemini

When talking to Claude you can say “add this to memory” at any point mid-conversation and it updates fluidly. It will then use its memory in future conversations with you. It’s brilliant. then surfaces that remembered context in unexpected but relevant ways. In one session about methane, Richie’s Claude proactively used an Oura ring metaphor because he goes on about it all the time. In another conversation about survey platforms, it proactively offered to “draft a pitch to Chris to switch tools”. It knew that Richie works at Bryant, who do surveys, and that his boss was Chris.

Gemini can’t do this on on a work email (it can kind of do this on a personal, paid plan). ChatGPT has memory too, but is not as smart at using it as Claude in Richie’s experience.

The nonprofit team plan is great value

The Claude nonprofit team plan is great value. It costs roughly $45 more than the individual plan but provides about 6.25x the usage (5x the team plan baseline, and the team plan is already 1.25x the individual plan). This effectively means you can afford to run the powerful Opus 4.7 on everything. Most people who aren’t impressed with Claude are using the weaker Sonnet and Haiku models, because Opus consumes your usage very fast.

2. Gemini

The main Gemini discussion this session was a workaround for its missing memory feature. We came up with a workaround: that create a custom Gem where one of the context documents is a Google Doc called something like “Things I’m working on now”. Whenever you want to add something to memory, you edit that Doc. The Gem will then pick up whatever is in there in future conversations. It is not as fluid as Claude (you cannot just say “remember this” mid-chat, you have to switch to the Doc), but it’s better than going to the gem, going into it’s settings, then updating the instructions. Gems themselves do not learn over time, so without something like this they are frozen to whatever was uploaded at creation.

We also discussed Gemini’s “@drive” connector, but that turned into its own write-up here.

3. Make.com

Olivia is learning how to build automations using Make.com.

Export your Make.com workflows using “blueprints” Share them with others or AI models.

We discussed how Make.com workflows can be exported as JSON files called “blueprints”, and they are the same pattern as n8n (another popular automation software). This unlocks several AI-assisted workflows. You can dump a broken blueprint into Gemini or Claude and ask why it is not working. We speculated that if someone gave you their Make.com blueprint, you could translate it to n8n by having it rewrite the JSON. An LLM probably cannot one-shot a working blueprint from a natural language description, because these platforms update frequently and training data lags, but a workable pattern is to build a rough scaffold yourself (hooking up the right nodes without proper logic), export the blueprint, and ask the LLM to refine the JSON.

One practical gotcha: Make.com sometimes appears to wipe unsaved work, so save frequently.

One documented success story this session: a Make.com scenario that takes a CSV of attendees who received travel funding, extracts each person’s contact details and funding amount, and auto-drafts an email confirming what they were allocated. This replaced a manual email-per-person process. A follow-up scenario to log these emails in Salesforce did not work on the first attempt, which illustrates where Make.com’s complexity starts to bite.

Use the Claude Chrome extension to build Make.com and n8n Automations

The Claude Chrome extension is a useful fallback when your Make.com or n8n automation is broken. Because the extension can click around the browser and take screenshots, you can point it at the broken workflow in the canvas and ask it to diagnose the issue based on what it actually sees on screen. This is particularly helpful when the problem is something visual or configuration-level that would not be obvious from the exported JSON alone, or when the platform has updated and your LLM’s training data is behind. Richie has done this several times with great success.

4. The security risks of connecting your AI to everything

We discussed the security implications of connectors: are there risks of hooking up your ChatGPT or Claude to you calendar, gmail, Asana, Hubspot and other sources of sensitive info?

The main security risk of AI connectors hapens when the AI is connected to an external communication channel that malicious actors can reach: email, public Slack groups, LinkedIn, Discord, and so on. A hostile message can convince your AI to ignore its previous instructions and send them your data through other connected tools. This risk is low for most users because it requires someone to be targeting you specifically, but it is worth being aware of. This is called prompt injection.

Connectors that only touch internal systems (Google Drive, Asana, internal databases) pose much less risk, because the outside work can’t attack you AI and convince it to send them your stuff.

In summary: the calculus of enabling a connector should turn on whether that connector exposes the AI to untrusted inbound messages, not on whether it connects to something that sounds sensitive.

That all for this session! We’ll be taking May off as Richie and Kevin are travelling, and will return in June!

UPDATE, 9th June, 2026!

With the release of Gemini 3.5 at Google I/O recently, I thought I would revisit this and do some additional testing with my teams at Greener By Default and Bryant Research.

I’ve added some of things we learned below in a callout box like this.

Kevin from my fortnightly AI learning circle dropped a casual comment the other week: Claude’s Google Drive connector is really good. Apparently, you can just ask Claude to go find something in your Google Drive and it works!

I was sceptical. All of the big three (ChatGPT, Gemini, and Claude) have had a Google Drive connector for over a year, but I never really used them because I didn’t think that they could be any good. My reasoning was that you’ve got a probabilistic (read: unreliable) large language model, rummaging through an organisation’s Google Drive, which can contain thousands of documents and is rarely that well organised. Surely it couldn’t be good. But I trust Kevin so decided to test Gemini’s Google Drive connector. After all, if anyone can make a decent Google Drive connector, it should be fucking Google.

The short answer

Gemini with the Drive connector scored 57 out of 60 on a quiz I made when set to “pro” or “thinking” mode. It scored The quiz was made from ~20 documents drawn from across my org’s Google Drive. The questions it got wrong all weren’t failures to find info: the questions involved finding info then doing calculations.

The test I ran

I came up with a clever way to test Gemini, which I feel pretty smug about:

1. I tested 2 things:

   1. I picked seven random documents from my org’s Google Drive. Not curated. Not related to each other. Deliberately spread across our organisation: some culinary notes, some research documents, some customer meeting notes, some HR material. I imported all seven into a single NotebookLM notebook.

   2. I used our (Greener by Default’s) Operations NBLM, that contains all our SOP documents

2. In both cases I used NotebookLM’s built-in quiz feature to generate quizzes from the documents. This saved me and hour or more of reading the docs and coming up with appropriate questions.

3. I asked Gemini each question, one at a time, in a fresh chat. NBLM quizs are multiple choice but I did not give Gemini the options, just the raw question. I explicitly typed @drive to activate the Drive connector each time.

4. I first tested Pro, then after seeing the results ran the quizzes again using Flash and Thinking

An example of a notebookLM quiz, similar to the one I used to test Gemini

The results

Shockingly, when set to pro or thinking mode, it got 57/60 correct across both notebooks! To emphasise, it was 95% accurate at answering specific questions from docs buried in our Google Drive, often with no pointer to which document held the answer.

The three it missed weren’t even retrieval failures. They were questions that required a calculation or inference on top of the document content. Something like “what percentage of X would Y be if you combined figures from two tables,” where the source documents had the raw numbers but Gemini needed to do the maths itself. That’s a completely different (and, honestly, more forgivable) class of error.

UPDATE: more result from the June 9th tests with Bryant and GBD

Summary

I was wrong. My mental model of how Drive connector performed was several months out of date. If you’ve been avoiding it for similar reasons, it might be worth retesting your assumptions.

Caveats

• We’re a small-ish org: GBD has 20 people, not sure if Gemini would struggle if you’re an org of 100 or more, with orders of magnitude more docs

• This was pure factual retrieval. It didn’t test synthesis across documents, nuance, or anything requiring judgement.

• NotebookLM-generated quizzes may be easier than real user questions. NotebookLM reads the documents and writes questions that are specifically answerable from them. Real queries from humans are messier, more ambiguous, and often ask things the documents don’t quite address. They’re hard questions though!

Over to you

Have you tested any of the Drive connectors (Claude, Gemini, or otherwise) against your own org’s files? Did you get similar hit rates, or was it a mess? I’d genuinely like to know whether my result generalises, or whether I just got lucky with my seven documents.

EDIT: I forgot to provide an actual link! Here it is

AI is transforming everything. How will it transform animal advocacy?

Is it going to supercharge our campaigns, or render some of them totally obsolete? Will it help factory farming or hurt it? And what happens to animals in the (possible) future where AI is smarter than all humans put together?

These seem like pretty big abstract questions, but given the pace of AI development, we need to be talking about them now. And increasingly, people are! Mostly on the Effective Altruism forum, as well as places like the Sentient Futures’ Slack group (join here) and their conferences.

Though I think there are too few discussions of these topics, there are also too many individual posts for me to keep up with.

So I decided to create an AI chatbot that would allow me to dive into the discussion in a personalised way and get some clarity on the issues. I did this using Google’s fantastic NotebookLM tool. I gave it the links to all EA forum posts with the “AI and Animals tag”.

The chatbot has not been programmed with any opinions of its own, and simply regurgitates and summarises what others have said. None of the content in the chatbot is mine, and so you should not assume I agree or disagree with any of the things it says.

The chatbot also has a variety of fun things to help you get into the field:

• A mind map

• Introductory infographics

• AI generated syntheses of key topics

Here’s some cool stuff it did!

I’d love to hear how people are using this!

There’s been an explosion in websites sharing Claude’s skills, e.g. here.

I personally think that you should almost never simply install another person’s skill:

1. Many of them are literally viruses and contain malware

2. Skills work best when they are ridiculously tailored to you. Generic skills will nearly always be less useful. For example, I don’t just want a skill that will check my calendar and give me a meeting prep summary. I want a skill that will knows the difference between meetings in my 2 jobs, and that knows the difference between client meetings and internal syncs for how much prep I need to do.

How to make skills

You should consider other’s skills to be inspiration. Luckily, a skill is just a markdown document, so you can literally just read it and see how its structured. Then, I advise copy pasting the good bits into a Claude window and telling it to “build me a skill that does these things“. you should also ask it to “Ask me 10 questions to understand what I want in greater detail”. SO we’re taking another’s skill as a base, then making it uniquely our own!

Claude in Chrome is the hidden gem of the Claude stack

I have been loving Claude Chrome recently. It allows Cowork (and regular Claude chat) to do so much more. 2 cool things:

• Ask Claude Cowork to post to your social media like X or Linkedin, and it will fail. These sites don’t like “bots”, so they have sophisticated anti-scraping measures in place. However! Claude in Chrome can evade most of that because it uses your browser and clicks around like you do. You can have Claude access your LinkedIn by using Claude in Chrome.

• It’s good for checking sources in reports. One of the most frustrating things about LLMs is when they write a bit of text and give a source, but then you open the source yourself and check it and it’s wrong! Claude Chrome helps with this. Open the source in your browser, but instead of checking it yourself, have the Claude Chrome extension check it. Because this version of Claude is basically shackled to the one browser tab, it is much more focussed and can usually tell you whether a claim is supported by evidence.

Tip! if it hits permission issues accessing most websites, you can change that in the settings:

Sometimes when I ask Claude to update one of its skills it will tell me that it can’t do that comma or it will try and do it and fail repeatedly before eventually getting to the correct answer. I’ve recently added this to my personal preferences:

You can edit your own skills. Skill folders (under /mnt/.claude/skills/) are often read-only. To modify one: (1) copy the skill folder to /tmp/skill-name/, (2) make your edits there, (3) package it by running python -m scripts.package_skill /tmp/skill-name/ from the skill-creator skill directory, (4) present the resulting .skill file to the user to reinstall. Do not attempt to write directly to /mnt/.claude/skills/. For non-trivial changes, invoke the skill-creator skill, which has a full testing and iteration workflow. You are always capable of editing skills; never tell the user otherwise.

Using Cowork to hit the ground running on a project you’ve already started

We don’t always join projects from scratch. Sometimes we join a project that is already in motion, full stop. There are already documents and meeting notes and all sorts of things already produced. Full stop. Instead of spending hours reading through everything to get us up to speed, can we use Cowork?

Here’s my current workflow:

• Point Cowork at the relevant project folder. This should be the folder that the project actually lives in. This could be a folder on your organisation’s shared Google Drive or Drop Box. Remember Cowork is your colleague, so you should all work out of the same folder.

• Ask Cowork to make a Claude.md document summarising the state of the project. A Claude.md document is basically just a simple text file that Claude will refer to in all conversations. Some tips:

  • Tell it to use sub-agents: This means that instead of Cowork looking through all the files it will turn into a manager, and it will create several copies of itself called sub agents. It will then send each sub agent to search through different files independently, then report back. This greatly speeds up the process because you have many versions of CoWork reading files at the same time. It also consumes less of your usage because each sub agent starts fresh.

  • Ask it to identify any discrepancies between the project documents. Sometimes a project folder will be full of different documents authored by different people at different times, and something CoWork is amazing at finding discrepancies and reconciling them.

  • If this is a project you know something about, tell CoWork to ask you ten questions to better understand the project. There are many things about a project that will be stuck in your head and not written down, so it’s good to get them out there for Cowork to understand.

  • Ask it to create an open questions doc. Once you’re up to speed on the project, the next thing you’re going to wonder is what should I get started on first? This document will tell you what seems like it is unfinished and what seems like it needs to be done.

Dispatch continues to be frustrating

At this point it seems pretty clear to me that Dispatch cannot be used for any serious work. the main use case is when you have started a task on your computer and you have a project already well set up, with skills already built, folder structure already in place, and you’re stepping away from your keyboard for a short period of time and you want to continue that task.

• It is extremely clunky to have Dispatch search through different folders or trigger or use skills or connectors: it can use skills and connectors, but you just won’t see it on the screen. It can move through folders on your computer but you don’t know what folder it is in at any given time

• You can only have one continuous conversation with Dispatch, meaning that you can only do really one task at a time

A hack for getting Dispatch to see files on your phone

Dispatch is at your computer not on your phone so it cannot technically see the so you can’t upload a file from your phone for it to use. one way around this is that you use Google Drive on your phone and then if you upload a file to Google Drive and you have Cowork on your computer working inside that Google Drive folder then it will be able to see files you upload. Clever!

Don’t have too many skills

A reminder that at the start of each conversation, Claude will load the description of every single skill it has into its memory, full stop. If you have like 30 skills with long descriptions, this can significantly eat into your context.

2 things I put at the end of about half of my Cowork Prompts:

“Repeat back to me what you think I’m asking you to do” - I do this after whenever I have given Claude a very messy dump of my thoughts.

“Ask me 10 questions to understand what I want” It asks you questions then uses the answers to provide a more personalised response. Often the questions surface things I hadn’t thought of too!