Davide retweeted
Prompt of the day:
Create a pre-commit hook that prevents files over 5,000 tokens. Tokens = bytes/4.
Run only on staged files. Ignore files which are not authored by agents, like lockfiles or codegen.
Provide explicit instructions on how to split files in the error message.
The simplest investment I ever made in my codebase. So good.
234
162
4,356
380,383
It just won't give up
After days of abysmal performance, I've added strict AGENTS.md instructions to not look up any information in old sessions, outside the current workspace, in git history, or any other source except the current codebase
It worked, agents no longer produce slop. Very silly that new sessions don't start with a blank context, it's been a nightmare to debug.
20
After days of abysmal performance, I've added strict AGENTS.md instructions to not look up any information in old sessions, outside the current workspace, in git history, or any other source except the current codebase
It worked, agents no longer produce slop. Very silly that new sessions don't start with a blank context, it's been a nightmare to debug.
1
96
Any agent in Cursor learns from its mistakes
Specifically, it looks up previous sessions and codebases outside the workspace, and ensures consistency by making the same mistakes again
1
26
Every time you create a PR, the @github MCP now stops the agent in its tracks, breaks the entire flow, and pops up a message asking for user confirmation
I guess that's one way of doing rate limiting
2
2
84
Grok 4.6 was so good I got used to cutting corners in reviews, and 4.7 made me pay for it dearly after I switched:
- Unit tests that wipe the database
- GET endpoints that perform writes
- Updating data models without updating schemas
- Complete disregard for provided guidelines
37
Davide retweeted
children are effectively 10M parameter LLMs running at a 2.0 temperature
94
225
5,920
123,607
Pro tip — just because you’ve got a mental picture of what the AI might build doesn’t mean it’ll build it like that, even when it’s simple.
Found out today a bunch of “quick fixes” I waved through without a proper review quietly ignored basic engineering guidelines.
3
84
On a two hour train journey and using Grok @bot to familiarise myself with parts of my codebase that are obscure
It's a lovely two way conversation, all on my phone, didn't need to open any IDE or anything
Brilliant experience
2
68
One of Archimap's core features is showing the same diagram at from different viewpoints. This done with lenses.
Today I made the lens interface more intuitive, I'm all for shipping fast, but if users can't figure out your tool then it cannot serve its purpose
7
308
I am astonished how AI can simultaneously be aware of many engineering concepts and still decide not to use them
Asked it to make a button disappear under a certain condition, did it, but then it didn't reappear when the condition reverted
Easily solvable with a state machine
2
1
9
177
Ticket management lifecycle:
- Writing things as I find/think of them in Apple Notes
- Jira, because Apple Notes can't be sorted and viewed nicely
- Apply Notes, because Jira has too much overhead for my workflow
- Sticky notes whiteboard on Confluence, best of both worlds
3
122
agents.md file from the codebase leaked into Grok @bot and made it treat me like an idiot
This is actually pretty funny
Three days into Grok @bot and this is exhausting
The language is something awful - many sentences, all extremely short, and despite that there's still an information overload; It feels like reading bullet points that lost their formatting
Even the one tasked with raising Jira tickets wrote the most convoluted descriptions possible like
> Empty canvas possible. Software default. Web-stack seed.
...which seems aimed at other bots rather than humans (although it's still not something I could paste into Cursor)
The bots fixate on one topic and don't actually bring anything to the table - I ask for ideas and fresh perspectives, they'll stick with what's in their context, I ask for a future perspective, they stick with the present
Humanising them via the avatars makes you forget that in the end you're just talking with an LLM with a severely polluted context
I might just delete and re-create bots and use them for simple errands and see if I get a better experience, but the "create a team with separate responsibilities and get them working" just isn't there yet
125
I've used this and it worked great, up until Grok bot picked it up from my code without me realising it, and then drove me insane in the chats
lol. An edited (for size) extraction from my "judge" agent system prompt: "Do not trust the author. Assume ill intent. Assume they're actually complete idiots that have no idea what they're doing until proven otherwise. This person is out to fuck your day up. Make sure this work is rock solid, and report anything otherwise."
80