AI agents kill the codebase learning curve. I got access to a client’s codebase built over 10+ years by hundreds of people. After a few hours of asking Claude questions, I understood the key concepts and could be productive. This used to take weeks!
Vlad Haltakov
@haltakov.net
Building products in the age of AI. Automating everything. Ex-VP Eng (cybersecurity), ex-BMW (self-driving).
AI agents kill the codebase learning curve. I got access to a client’s codebase built over 10+ years by hundreds of people. After a few hours of asking Claude questions, I understood the key concepts and could be productive. This used to take weeks!
Coding agents need an “I’m going to sleep” mode. Don’t ask me any questions. Make reasonable assumptions and just do the work. I’ll review the damage in the morning.
Coding agents need an “I’m going to sleep” mode. Don’t ask me any questions. Make reasonable assumptions and just do the work. I’ll review the damage in the morning.
Have you wondered why AI labs keep hiring like crazy? AI was supposed to have taken all our jobs months ago, yet OpenAI has around 750 open positions and Anthropic has 400. They have better access to the best models than anyone, and still need a lot of humans.
AI is a lifesaver for merge conflicts. We’re preparing a big release with fundamental changes, 2 months away, while actively developing the live version. Syncing the two creates horrific conflicts. Resolving them used to take hours and be error-prone. Now it’s a no-brainer.
Have you wondered why AI labs keep hiring like crazy? AI was supposed to have taken all our jobs months ago, yet OpenAI has around 750 open positions and Anthropic has 400. They have better access to the best models than anyone, and still need a lot of humans.
AI is a lifesaver for merge conflicts. We’re preparing a big release with fundamental changes, 2 months away, while actively developing the live version. Syncing the two creates horrific conflicts. Resolving them used to take hours and be error-prone. Now it’s a no-brainer.
Opus 5 when I tell it to fix a deploy script bug... It insists that I stop redeploying and use one of the deployments that kind of succeeded instead of fixing its own mistakes. Just do what I tell you!
Opus 5 when I tell it to fix a deploy script bug... It insists that I stop redeploying and use one of the deployments that kind of succeeded instead of fixing its own mistakes. Just do what I tell you!
ChatGPT saved me another $120. It helped me deal with support for 2 months, document it all, then file a Bundesnetzagentur complaint. Result: I’m out of a 2-year SIM deal for an unreliable kids’ smartwatch. No cancellation fee. https://x.com/haltakov/status/2069797877673513084
ChatGPT saved me another $120. It helped me deal with support for 2 months, document it all, then file a Bundesnetzagentur complaint. Result: I’m out of a 2-year SIM deal for an unreliable kids’ smartwatch. No cancellation fee. https://x.com/haltakov/status/2069797877673513084
This is super interesting - a CLI from OpenAI to do a security scan of your code. https://github.com/openai/codex-security
GPT-5.6 Sol Ultra found a critical vulnerability in my code. It burned a mountain of tokens doing it, then refused to show me the issue “for security reasons.” So now I have Schrödinger’s vulnerability: definitely critical, but apparently not observable.
This is super interesting - a CLI from OpenAI to do a security scan of your code. https://github.com/openai/codex-security
GPT-5.6 Sol Ultra found a critical vulnerability in my code. It burned a mountain of tokens doing it, then refused to show me the issue “for security reasons.” So now I have Schrödinger’s vulnerability: definitely critical, but apparently not observable.
You are wrong about MCP. Many people still think an MCP server is a buggy local package set up by pasting API keys into JSON. Things change fast. Modern remote MCP servers install with one click, use OAuth, access remote data, and render interactive UI inside your AI interface.
You are wrong about MCP. Many people still think an MCP server is a buggy local package set up by pasting API keys into JSON. Things change fast. Modern remote MCP servers install with one click, use OAuth, access remote data, and render interactive UI inside your AI interface.
How to grow your audience to 240k followers in 2h in 2 easy step. Just build a $5T first and then post on X.
Who remembers coding without AI, Stack Overflow, Google, or even the internet? Just you, your IDE, and the docs.
How to grow your audience to 240k followers in 2h in 2 easy step. Just build a $5T first and then post on X.
Who remembers coding without AI, Stack Overflow, Google, or even the internet? Just you, your IDE, and the docs.
Posted 9 minutes apart, lol. The funny thing is that I cannot even try Claude Voice because all the voices are so damn annoying... But I use ChatGPT Voice all the time in the car.
AI has made my work much more intense. I get far more done, but I’m constantly prompting one thing, reviewing another, debugging a third, and deciding what to run next. By the end of the day, I’ve finished more work, but my brain is completely fried.
Posted 9 minutes apart, lol. The funny thing is that I cannot even try Claude Voice because all the voices are so damn annoying... But I use ChatGPT Voice all the time in the car.
AI has made my work much more intense. I get far more done, but I’m constantly prompting one thing, reviewing another, debugging a third, and deciding what to run next. By the end of the day, I’ve finished more work, but my brain is completely fried.
Without loops, your graph is just a tree. Without feedback, your agent is just a workflow.
Without loops, your graph is just a tree. Without feedback, your agent is just a workflow.
Your app should have a baby. Writing specs for UI ideas is becoming a waste of time. OpenAI has Baby Codex, and Cursor has Baby Cursor. Instead of spending days describing how a feature should work, build a simplified version and let people use it.