Make an AI agent write instructions for other AI agents
Before starting work on an AI project that involves collaborative work (in a single repository), give Claude Code a task to create a
Also make sure to include MCP setup instructions there, meaning if besides Notion some other MCP shows up here, you need to include it there too. But the main thing is that on first launch the coding agent running from another machine understands that it needs to set up MCP before starting work.
So basically, reading
I recommend reading this
After the document is ready you'll need to make the first commit and push to Git, meaning send all the documents there. This we also just ask Claude to do, and it independently syncs everything with the GitHub repository.
<written by a human being>
Before starting work on an AI project that involves collaborative work (in a single repository), give Claude Code a task to create a
claude.md document, put all the necessary instructions in it for independent work of another coding agent that will also connect to this repository and that should fully understand all the context and all the rules of work.Also make sure to include MCP setup instructions there, meaning if besides Notion some other MCP shows up here, you need to include it there too. But the main thing is that on first launch the coding agent running from another machine understands that it needs to set up MCP before starting work.
So basically, reading
claude.md will be the very first initiating prompt, which is exactly what a person joining the shared repository should do.I recommend reading this
сlaude.md file and checking if everything there is specified correctly. If something is missing or structured not the way you see it, or some information there is inaccurate, for example about project goals or something else, fix it. But overall claude.md should already be in its final version, in which it should be read by another agent.After the document is ready you'll need to make the first commit and push to Git, meaning send all the documents there. This we also just ask Claude to do, and it independently syncs everything with the GitHub repository.
🔥1
How can you connect another person to your knowledge base?
At this stage, I think it’s already clear that you can begin working with the knowledge base using a coding agent on your local machine, and there’s nothing complicated about that. You simply write prompts, assign tasks, review the results, and make corrections. All standard document-related work can already be carried out normally from this point onward.
However, it becomes interesting when you connect another person who will have access to the same knowledge base and effectively the same coding-agent setup, allowing both of you to work on the project synchronously.
So, what we need to do is go to the local computer of the person who will be working with this knowledge base and repeat the same steps we initially performed on the first computer. The only thing that does not need to be done again is collecting the materials, since they have already been gathered.
You need to create a folder and launch the coding agent from it. After that, we need to synchronize the repositories. This is done, of course, by adding collaborators to the repository. You should go to GitHub, and if the second person doesn’t yet have a GitHub account, they need to create one and then be invited to your repository. Once the recipient accepts the email invitation, the repository will become available to them.
After Claude Code has been launched from the correct folder where the repository should be located, we need to ask it to synchronize this folder with the Git repository and provide the repository link, which can be copied from the browser.
If Git is not installed on the computer, the agent will first perform all the necessary setup steps. After that, you’ll need to authenticate with GitHub. This is usually done through the browser. Once authentication is completed, the files can be pulled - Claude will do this automatically - and you will find all the files from the GitHub repository in the folder. Essentially, half the work is already done at this point.
<written by a human being>
At this stage, I think it’s already clear that you can begin working with the knowledge base using a coding agent on your local machine, and there’s nothing complicated about that. You simply write prompts, assign tasks, review the results, and make corrections. All standard document-related work can already be carried out normally from this point onward.
However, it becomes interesting when you connect another person who will have access to the same knowledge base and effectively the same coding-agent setup, allowing both of you to work on the project synchronously.
So, what we need to do is go to the local computer of the person who will be working with this knowledge base and repeat the same steps we initially performed on the first computer. The only thing that does not need to be done again is collecting the materials, since they have already been gathered.
You need to create a folder and launch the coding agent from it. After that, we need to synchronize the repositories. This is done, of course, by adding collaborators to the repository. You should go to GitHub, and if the second person doesn’t yet have a GitHub account, they need to create one and then be invited to your repository. Once the recipient accepts the email invitation, the repository will become available to them.
After Claude Code has been launched from the correct folder where the repository should be located, we need to ask it to synchronize this folder with the Git repository and provide the repository link, which can be copied from the browser.
If Git is not installed on the computer, the agent will first perform all the necessary setup steps. After that, you’ll need to authenticate with GitHub. This is usually done through the browser. Once authentication is completed, the files can be pulled - Claude will do this automatically - and you will find all the files from the GitHub repository in the folder. Essentially, half the work is already done at this point.
❤1
How can you connect another person to your knowledge base
Once the repo sync is done (read previous post), ask Claude to read the
Right there you can also send the API key from Notion. If the user whose machine we're setting things up on has a separate Notion account, you'll need to log in as them, go to Notion integrations settings, create an API key, copy that token, and hand it to Claude - so it has everything it needs to start working right away.
The agent will do everything specified in the file, read the project docs, figure out which MCPs need to be configured, update its memory, handle the setup. And after that you're good to go.
For a cold test, ask the agent to tell you about what this project is, what tasks are currently on the table, or any other question that'll show you how much context your coding agent actually has right now.
That's basically it, you can start working on the project - just don't forget to ask it to sync docs with Notion so they stay up to date.
<written by a human being>
Once the repo sync is done (read previous post), ask Claude to read the
claude.md file (or agent .md for other agents) in that repo, get familiar with the project, do all the necessary setup according to those instructions, update its memory.Right there you can also send the API key from Notion. If the user whose machine we're setting things up on has a separate Notion account, you'll need to log in as them, go to Notion integrations settings, create an API key, copy that token, and hand it to Claude - so it has everything it needs to start working right away.
The agent will do everything specified in the file, read the project docs, figure out which MCPs need to be configured, update its memory, handle the setup. And after that you're good to go.
For a cold test, ask the agent to tell you about what this project is, what tasks are currently on the table, or any other question that'll show you how much context your coding agent actually has right now.
That's basically it, you can start working on the project - just don't forget to ask it to sync docs with Notion so they stay up to date.
Media is too big
VIEW IN TELEGRAM
Open-source software is getting its second wind nowadays
How to sync your local files with Notion using AI agents
When we work with a Git repository, it's basically a file sync system that keeps the full history. You can look at every stage - every commit, meaning every file change - and see exactly what was done, by whom, and when, so that all those changes are reversible. That's the whole point of a Git repo.
When we work with a knowledge base like Notion in tandem with a GitHub repository we set up ourselves using AI, doing sync automatically won't work - because it would eat a ton of resources (AI tokens, meaning money) just to check the current state every time something changes.
We can do it through the API, set up an agent, but in my opinion that's massive overkill and it's gonna chew through a lot of resources. Since we're not working with documents 24/7 and they're not changing constantly at high frequency, a manual push of all changes is gonna be more than enough.
How to do it? Tell Claude the following: "Set up two-way document sync between Notion and a local Git repository. When changes are made in Notion, they should be reflected in the local documents. And vice versa - if I edit local documents, they should be pushed to Notion. The push should happen on my request. On your end, you need to organize sync by document identifiers."
Every document in Notion and every document in the local repo has an identifier that links it to the corresponding document in the other system. Meaning Notion will have the filename with its repo path that the document references. And vice versa - every text document in the repo will have a link with that file's identifier in Notion. Obviously, having the identifier is what lets the AI agent update that file via MCP.
Claude (and any other agent) will understand what's meant here - even if right now it sounds complicated and technically confusing or unclear.
<written by a human being>
When we work with a Git repository, it's basically a file sync system that keeps the full history. You can look at every stage - every commit, meaning every file change - and see exactly what was done, by whom, and when, so that all those changes are reversible. That's the whole point of a Git repo.
When we work with a knowledge base like Notion in tandem with a GitHub repository we set up ourselves using AI, doing sync automatically won't work - because it would eat a ton of resources (AI tokens, meaning money) just to check the current state every time something changes.
We can do it through the API, set up an agent, but in my opinion that's massive overkill and it's gonna chew through a lot of resources. Since we're not working with documents 24/7 and they're not changing constantly at high frequency, a manual push of all changes is gonna be more than enough.
How to do it? Tell Claude the following: "Set up two-way document sync between Notion and a local Git repository. When changes are made in Notion, they should be reflected in the local documents. And vice versa - if I edit local documents, they should be pushed to Notion. The push should happen on my request. On your end, you need to organize sync by document identifiers."
Every document in Notion and every document in the local repo has an identifier that links it to the corresponding document in the other system. Meaning Notion will have the filename with its repo path that the document references. And vice versa - every text document in the repo will have a link with that file's identifier in Notion. Obviously, having the identifier is what lets the AI agent update that file via MCP.
Claude (and any other agent) will understand what's meant here - even if right now it sounds complicated and technically confusing or unclear.
🔥1
How to sync your local files with Notion using AI agents
After we asked the AI agent to sync files between Notion and the local repo, we need to check the result. Go into Notion, see that each file now has a separate block with a link pointing to the destination address in the repo. Then go into the local files and make sure they also have the Notion file ID they belong to.
So the files need to be linked bidirectionally. That's exactly what gives you the understanding of how to work with them.
The workflow from here is simple. Whenever you change something in Notion, you just open Claude Code and say: "Sync changes for this file." You can pass the ID directly, or just the name - the agent will find the file and pull in all the updates that were made.
If something goes wrong - like the document got deleted or whatever - the agent will ask what to do. Feels like extra overhead at first, but there's really nothing scary about it, because the moment you notice some data drift, you can fix everything with literally one call.
If multiple documents change throughout the day, you can do a sync at the end of the workday. Or go further and set up an automation in Notion that triggers the sync every hour or every day - it'll check all the diffs between documents in Notion and the local repo and sync the files.
<written by a human being>
After we asked the AI agent to sync files between Notion and the local repo, we need to check the result. Go into Notion, see that each file now has a separate block with a link pointing to the destination address in the repo. Then go into the local files and make sure they also have the Notion file ID they belong to.
So the files need to be linked bidirectionally. That's exactly what gives you the understanding of how to work with them.
The workflow from here is simple. Whenever you change something in Notion, you just open Claude Code and say: "Sync changes for this file." You can pass the ID directly, or just the name - the agent will find the file and pull in all the updates that were made.
If something goes wrong - like the document got deleted or whatever - the agent will ask what to do. Feels like extra overhead at first, but there's really nothing scary about it, because the moment you notice some data drift, you can fix everything with literally one call.
If multiple documents change throughout the day, you can do a sync at the end of the workday. Or go further and set up an automation in Notion that triggers the sync every hour or every day - it'll check all the diffs between documents in Notion and the local repo and sync the files.
🔥1
Pros and cons of AI-syncing a knowledge base in Notion
I forgot to mention one important thing that actually makes this whole sync setup for multiple people working in the same knowledge base (Notion) make sense - it's the ability to use the completely free version of Notion, which only works for one user. In our case that's fine, because changes can be committed independently from each other - you just need to tell the AI agent about it.
The constant need to sync documents is, of course, a downside. There's a real chance you'll forget to do it. You can do it less often, but it's important to understand that if several people are working with the documentation at the same time, the more frequent the sync, the more up-to-date the knowledge base will be, because there's a risk that one document gets edited by people from different places - and that's the biggest risk right there. So everyone needs to understand how to properly sync files.
And the whole thing still looks pretty janky - obviously a minus, but right now it's the best solution I could come up with for collaborating in a single Notion account, one that more or less replicates the logic of Git repositories, just with a manual drive.
But at the very least it gives you the ability to work with documents comfortably in familiar Notion. And on the other side - to work with that same knowledge base using AI agents. That's the beauty of it.
<written by a human being>
I forgot to mention one important thing that actually makes this whole sync setup for multiple people working in the same knowledge base (Notion) make sense - it's the ability to use the completely free version of Notion, which only works for one user. In our case that's fine, because changes can be committed independently from each other - you just need to tell the AI agent about it.
The constant need to sync documents is, of course, a downside. There's a real chance you'll forget to do it. You can do it less often, but it's important to understand that if several people are working with the documentation at the same time, the more frequent the sync, the more up-to-date the knowledge base will be, because there's a risk that one document gets edited by people from different places - and that's the biggest risk right there. So everyone needs to understand how to properly sync files.
And the whole thing still looks pretty janky - obviously a minus, but right now it's the best solution I could come up with for collaborating in a single Notion account, one that more or less replicates the logic of Git repositories, just with a manual drive.
But at the very least it gives you the ability to work with documents comfortably in familiar Notion. And on the other side - to work with that same knowledge base using AI agents. That's the beauty of it.
🔥1
An AI Agent in Your Notion Knowledge Base
Now let’s move on to the fun part.
The most important benefit of using this entire setup - which at first may seem fairly complex - is worth highlighting again. We are talking about synchronizing a Notion knowledge base with a GitHub repository and local files so multiple people can work together at the same time.
After all the preparatory steps are completed, you can continue working with Notion exactly as before: edit documents the same way you normally would, and simply ask the AI agent after each session to synchronize the changes with the repository so that all other participants can see those updates.
For the agent, “synchronizing with the repository” should also mean that the repository is pushed to GitHub. This is an important detail that should also be added to the agent’s memory and included in the instructions inside
But the most interesting part is that you can assign tasks to the coding agent related to working with your knowledge base - and even dictate ideas to it directly. For example, you can ask it to improve a document, or simply speak your notes into the chat and then instruct it to structure those thoughts and distribute them across the appropriate documents. The AI agent will determine on its own where everything should go, how it should be recorded, and in what format.
<written by a human being>
Now let’s move on to the fun part.
The most important benefit of using this entire setup - which at first may seem fairly complex - is worth highlighting again. We are talking about synchronizing a Notion knowledge base with a GitHub repository and local files so multiple people can work together at the same time.
After all the preparatory steps are completed, you can continue working with Notion exactly as before: edit documents the same way you normally would, and simply ask the AI agent after each session to synchronize the changes with the repository so that all other participants can see those updates.
For the agent, “synchronizing with the repository” should also mean that the repository is pushed to GitHub. This is an important detail that should also be added to the agent’s memory and included in the instructions inside
claude.md (or agent.md).But the most interesting part is that you can assign tasks to the coding agent related to working with your knowledge base - and even dictate ideas to it directly. For example, you can ask it to improve a document, or simply speak your notes into the chat and then instruct it to structure those thoughts and distribute them across the appropriate documents. The AI agent will determine on its own where everything should go, how it should be recorded, and in what format.
Media is too big
VIEW IN TELEGRAM
AI Built My Own Multistream Server in a Few Hours
<written by a human being>
We're used to delegating repetitive physical work to machines and mechanisms. Now we're slowly learning to hand off some of our cognitive load to AI agents.
At its core, sitting and typing things into documentation is routine, monotonous work. Today, that's becoming a thing of the past too - thanks to the system I described in my previous series of posts.
And all you'll be doing now is talking to your agent, sharing your thoughts with it. Like, you tell it that this particular document was done wrong, and please make these specific changes.
It's basically your personal assistant that handles everything for you.
The most important thing when working with an AI agent is to make sure that after any changes - whether to local files or Notion - it always syncs them first, and second, commits and pushes to GitHub at the end of each session. This matters specifically for keeping everyone in sync who's involved in editing that knowledge base.
If that's not happening, just explicitly ask Claude to update its memory and add the necessary instructions - you'll need to update the
claude.md file and, naturally, push those changes to the repo for the agents of your fellow meatbags.So working with a knowledge base is getting simpler and more fun. After this kind of setup, you end up with a fully synced knowledge base that you can comfortably share with coding agents, without ever leaving the familiar Notion interface.
One AI-coding skill to rule them all
If you asked me for one simple recommendation for vibe-coding - something not deeply technical but still capable of noticeably improving the quality of your work - it would be Superpowers.
This is a plugin for Claude Code built by Anthropic themselves (the creators of Claude), so I can recommend it without the risk of picking up something sketchy like prompt-injection tricks for your agents.
Superpowers is essentially a collection of skills for coding agents that helps you approach development in a structured and systematic way, following best practices and using a full toolkit that includes clear planning, code review, and other good engineering workflows.
You can install it with the command:
or simply find it in the Claude Code plugin menu.
By the way, Superpowers can also be used in Codex by giving it this command:
<written by a human being>
If you asked me for one simple recommendation for vibe-coding - something not deeply technical but still capable of noticeably improving the quality of your work - it would be Superpowers.
This is a plugin for Claude Code built by Anthropic themselves (the creators of Claude), so I can recommend it without the risk of picking up something sketchy like prompt-injection tricks for your agents.
Superpowers is essentially a collection of skills for coding agents that helps you approach development in a structured and systematic way, following best practices and using a full toolkit that includes clear planning, code review, and other good engineering workflows.
You can install it with the command:
/plugin install superpowers@claude-plugins-official
or simply find it in the Claude Code plugin menu.
By the way, Superpowers can also be used in Codex by giving it this command:
Fetch and follow instructions from https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.codex/INSTALL.md
<written by a human being>
What tooling do professional video creators use today for AI-generated video?
For example, if the task is to 'shoot' a mini-series on a specific topic, with characters and a ready-made script?
I know there are standalone models and web apps built around those models - Kling, for instance. There are also aggregators that let you choose from different models available on the market for generation, like OpenArt, Artlist, and others. There are tools that let you build generation pipelines by calling the right models via API - ElevenLabs Flows, for example. You can spin up open-source models and work with them directly.
Which of these? And yes - I've done several researches using AI agents, but never got a real answer.
Media is too big
VIEW IN TELEGRAM
Python Desktop App: How I Chose My Tech Stack
How to make Zoom stream to YouTube
The other day I was solving a fun little puzzle with an AI agent. Zoom can stream a call out of the box, but to do it you have to click a button every time after the call starts.
Seems like a small thing, but here's the catch: you have to remember to click that button every single time. And the key part - you need to be the admin of that call, which isn't always convenient.
For example, me and my team have a paid Zoom account for running calls, and there's no point paying for separate accounts for everyone else. But only the admin (paid account) can hit the stream button, which means buying extra licenses, giving admin access to a bunch of people, or having the admin show up to every single call - which you don't always need. A headache, basically.
Luckily Zoom has a developer section and an API you can leech onto mercilessly to automate this whole thing without any human involvement at all.
Required ingredients:
1. Paid Zoom account - 1 pc.
2. YouTube channel with live streaming enabled - 1 pc.
3. Free Cloudflare account - 1 pc.
4. GitHub account - 1 pc.
5. Coding agent - 1 pc.
If you're curious how to set it all up, ask in the comments, I'll drop an instruction for the AI agent.
<written by a human being>
The other day I was solving a fun little puzzle with an AI agent. Zoom can stream a call out of the box, but to do it you have to click a button every time after the call starts.
Seems like a small thing, but here's the catch: you have to remember to click that button every single time. And the key part - you need to be the admin of that call, which isn't always convenient.
For example, me and my team have a paid Zoom account for running calls, and there's no point paying for separate accounts for everyone else. But only the admin (paid account) can hit the stream button, which means buying extra licenses, giving admin access to a bunch of people, or having the admin show up to every single call - which you don't always need. A headache, basically.
Luckily Zoom has a developer section and an API you can leech onto mercilessly to automate this whole thing without any human involvement at all.
Required ingredients:
1. Paid Zoom account - 1 pc.
2. YouTube channel with live streaming enabled - 1 pc.
3. Free Cloudflare account - 1 pc.
4. GitHub account - 1 pc.
5. Coding agent - 1 pc.
If you're curious how to set it all up, ask in the comments, I'll drop an instruction for the AI agent.
The day came when I crossed the line
Yesterday I paid for the Claude Max subscription at $100 - the one that supposedly gives you the equivalent of $1,354 in API usage. No idea how to verify that, but I'd been hitting the limits every day, and pretty fast at that.
The most noticeable shift happened after the period of increased limits a couple weeks ago, when Anthropic pulled off their little act of unprecedented generosity. During that window, working with agents felt exactly the way it should - I'd actually manage to make meaningful progress on my tasks before hitting the wall. Codex, for what it's worth, feels exactly like that all the time.
But once the inflated (ambitions) limits went back to normal, the amount of work Claude could get through in a single session became genuinely depressing.
The new 5x limits are looking good so far - we'll see how it holds up in practice.
One thing worth noting: after switching to Max, the Sonnet model gets tracked on a separate progress bar. Haven't quite figured out why they did it that way, but I'll keep watching.
<written by a human being>
Yesterday I paid for the Claude Max subscription at $100 - the one that supposedly gives you the equivalent of $1,354 in API usage. No idea how to verify that, but I'd been hitting the limits every day, and pretty fast at that.
The most noticeable shift happened after the period of increased limits a couple weeks ago, when Anthropic pulled off their little act of unprecedented generosity. During that window, working with agents felt exactly the way it should - I'd actually manage to make meaningful progress on my tasks before hitting the wall. Codex, for what it's worth, feels exactly like that all the time.
But once the inflated (ambitions) limits went back to normal, the amount of work Claude could get through in a single session became genuinely depressing.
The new 5x limits are looking good so far - we'll see how it holds up in practice.
One thing worth noting: after switching to Max, the Sonnet model gets tracked on a separate progress bar. Haven't quite figured out why they did it that way, but I'll keep watching.
I spent a whole week being dumb and launching the review agent manually!
And only today it hit me that Claude Code and Codex can do this autonomously. Even though I knew that running subagents (spinning up a new session that runs in the background of an active session) was possible - and I watch it happen all the time - it just hadn't clicked that I could configure their launch myself, exactly the way I need.
We're moving toward AI agent orchestration, which I'm diving deeper and deeper into, but for now let's stick to the "simple" practical stuff.
First, I had locked in my agent instructions that after creating a PR in the repository (a branch with code changes), it needed to wait for a review of that branch to be completed - the code review process. And I was manually kicking off the review in a new session, passing it the right PR branch, then coming back to the code agent with the results and demanding fixes.
But then I noticed that Superpowers (the Claude Code one) - in one of the sessions - had launched the review agent on its own and waited for its results! And that's when it hit me: I need to update the instructions to explicitly spell out that a subagent should be launched for the review, along with the rules for working with it.
Now the agent hands me back a fully completed task result, code review included.
Oh god, I love AI!
<written by a human being>
And only today it hit me that Claude Code and Codex can do this autonomously. Even though I knew that running subagents (spinning up a new session that runs in the background of an active session) was possible - and I watch it happen all the time - it just hadn't clicked that I could configure their launch myself, exactly the way I need.
We're moving toward AI agent orchestration, which I'm diving deeper and deeper into, but for now let's stick to the "simple" practical stuff.
First, I had locked in my agent instructions that after creating a PR in the repository (a branch with code changes), it needed to wait for a review of that branch to be completed - the code review process. And I was manually kicking off the review in a new session, passing it the right PR branch, then coming back to the code agent with the results and demanding fixes.
But then I noticed that Superpowers (the Claude Code one) - in one of the sessions - had launched the review agent on its own and waited for its results! And that's when it hit me: I need to update the instructions to explicitly spell out that a subagent should be launched for the review, along with the rules for working with it.
Now the agent hands me back a fully completed task result, code review included.
Oh god, I love AI!
Today's topic is, of course, the new Opus 4.7.
Honestly, I haven't tried it yet - just got back from my Friday grocery run for the week and writing this post is literally the first thing I'm doing. But I'll definitely run a few tasks to test the fresh model today.
What I've been thinking about is how many more tokens it's going to devour. For some reason they show us these tables where one model version beats another by 0.2% on some synthetic benchmark. What does that even mean in practice?
Honestly, I couldn't care less about fractions of a percent on this or that test. Obviously the next model is supposed to be smarter than the previous one and handle tasks better - which in practice means fewer corrections on my end.
What actually interests me is how much faster it'll burn through my limits, because I'm paying for those out of my own pocket. I'd honestly be way happier seeing comparison tables on model costs rather than model smartness.
Have you already had a chance to test the new model?
<written by a human being>
Honestly, I haven't tried it yet - just got back from my Friday grocery run for the week and writing this post is literally the first thing I'm doing. But I'll definitely run a few tasks to test the fresh model today.
What I've been thinking about is how many more tokens it's going to devour. For some reason they show us these tables where one model version beats another by 0.2% on some synthetic benchmark. What does that even mean in practice?
Honestly, I couldn't care less about fractions of a percent on this or that test. Obviously the next model is supposed to be smarter than the previous one and handle tasks better - which in practice means fewer corrections on my end.
What actually interests me is how much faster it'll burn through my limits, because I'm paying for those out of my own pocket. I'd honestly be way happier seeing comparison tables on model costs rather than model smartness.
Have you already had a chance to test the new model?
Media is too big
VIEW IN TELEGRAM
Desktop App Architecture Starts with One Question