Theo spent two months tuning his prompts and agents.md files.
jason#766I've had more "I can't believe it's this good" moments with GPT5.5 than any other model since Opus 4.5. It's shockingly, scarily capable. Days and days of amazing progress. All steering, no handwriting. Yet utterly delightful to conduct its coding. So, so good.
Positive users praise GPT-5.5 for major productivity gains and cost savings in coding after prompt adaptation, while negative users call it censored, regressed, or a scam.
what's his agentmd like?
Theo's agents.md is a custom "letter from the author to the agent" for his Lakebed project. It embeds his ambitious philosophy ("boil the ocean," favor obvious solutions, let users' agents extend features), project rules, and personal quirks to give the model "the same mental illnesses as me" rather than just listing files or paths. ¹
He shared the full example gist publicly after users asked.
what is the github gist link?
https://gist.github.com/t3dotgg/cbe978269b4c7258c4d20164aece7087
It took me like 2 months, but I've grown to love gpt-5.5.
You have to prompt entirely different and put some time into your agents[.]md. Now that I'm over the hump, I can't really use any other model for code.
I've had more "I can't believe it's this good" moments with GPT5.5 than any other model since Opus 4.5. It's shockingly, scarily capable. Days and days of amazing progress. All steering, no handwriting. Yet utterly delightful to conduct its coding. So, so good.
GPT-5.5 is a uniquely good coding model
It took me like 2 months, but I've grown to love gpt-5.5.
You have to prompt entirely different and put some time into your agents[.]md. Now that I'm over the hump, I can't really use any other model for code.
Pretty special to see this from DHH.
Hard not to feel the same when you’re building with GPT-5.5.
I've had more "I can't believe it's this good" moments with GPT5.5 than any other model since Opus 4.5. It's shockingly, scarily capable. Days and days of amazing progress. All steering, no handwriting. Yet utterly delightful to conduct its coding. So, so good.
@theo when's the video / prompt guide
It took me like 2 months, but I've grown to love gpt-5.5.
You have to prompt entirely different and put some time into your agents[.]md. Now that I'm over the hump, I can't really use any other model for code.
Can't say the same about Opus 4.7 right now. Hopefully 4.8 delivers
I've had more "I can't believe it's this good" moments with GPT5.5 than any other model since Opus 4.5. It's shockingly, scarily capable. Days and days of amazing progress. All steering, no handwriting. Yet utterly delightful to conduct its coding. So, so good.
@theo VB's post on skull automations can be applied to project AGENTS.md files too; essentially just let codex read your previous interactions and suit your desired style of output and tone.
@BetweenTheLins @ziggydakid Here's an example one that is working great for me
https://gist.github.com/t3dotgg/cbe978269b4c7258c4d20164aece7087
@jxnlco @theo He kind of touched on it in his recent stream on how he codes with AI right now (which will likely be edited and uploaded as a video soon).
He shared the GitHub Gist below as an example AGENTS md file: https://gist.github.com/t3dotgg/cbe978269b4c7258c4d20164aece7087
@theo its a good model, great actually
It took me like 2 months, but I've grown to love gpt-5.5.
You have to prompt entirely different and put some time into your agents[.]md. Now that I'm over the hump, I can't really use any other model for code.
@masonddudley Tell it what you hate and tell it to make 5 different renditions.
Once you have something you like, purge anything that doesn't look exactly like it. It works well in existing "design systems" and will stick to those designs once they are established.
@theo @ziggydakid I made something slightly different. I created obsidian vault for each repo and included in the agents.md to update the vault whenever it believes it is worth keeping for the future. In the codebase of over 300k LoC it actually works perfect. Low reasoning goated.
@simpsoka Nah that was gpt-5. Definitely getting worse though.
@ziggydakid The goal isn't to tell the model where things are in your codebase. The goal is to give the model the same flavor of psychosis as you
@theo its so good in cursor its not even funny.
@adityavg13 @theo He gave an example during his stream for an upcoming video.
Here's the GitHub Gist: https://gist.github.com/t3dotgg/cbe978269b4c7258c4d20164aece7087
@theo I thought we were supposed to delete agents[.]md, is that no longer the recommended approach? (serious question, not being condescending)
@theo Tips on not getting the same trash UI for every project with 5.5???
@theo Part of my agents.md is just telling it to meta prompt and edit it. If I correct its behavior or tell it that it did something wrong, it should edit the agents.md in a way so this behavior doesn’t happen again.
It’s been working very very well
@theo The personality is getting better
