Cognitive resolution talks
Hi there. and thanks for the invitation
I'm looking for help about the best techniques to keep my AI cognitive resolution 100% safe for a "little" project I have in Cursor (~50k code lines).
READ THIS BEFORE SENDING ANY ANSWER:
First of all: I'm not a huge fan of cursorrules 1000 words files -> I consider that overwhelming AI with auto restriction goes against cognitive resolution since it contaminate the context window, and I'd kindly ask to those of you who think differently about it to ignore my post and to NOT REPSOND.
Ok now:
(Context) My rules are quiet simples:
- File length limited under 300 lines
- File nomenclature with metadatas (dependencies etc)
- Strict cone of files reading under 5.
- Composer 2.5 alerts me if my instruction is too complex and if I need to switch to a higher reasoning model
My 3 questions are:
- I wonder if the "Diff-Anchoring" technics (AI can select the lines to focus on in the instruction-concerned files) is truely effective?
- I'm thinking about the fact the agent may selects some lines at the very begining, and then after some on-going reasoning, it may loose the whole point of the instruction just because the file-context isn't complete anymore... what do you think about it?
The 5th rule I'm working on is:## STRICT LINE-SELECTION & DIFF-ANCHORING- NEVER read or rewrite a full file if only a subset of lines needs modification.- When referencing code, extract only the specific code-block or function using the \@ symbol or line-range (e.g., \@file.ts:120-150`).- Always output code modifications as localized diff blocks. Do not touch or hallucinate lines outside the requested block.``
- Directly connected to first rule - does the "State-Sync & Hard Reload" technic (a file reloading rule every 5 inner-reasoning-msgs) can really helps the Diff-Anchoring tecnic to keep the actions "bugs-free"
the eventual 6th rule which could complete the 5th one:## ANTI-REGRESSION & FILE RELOADING (Rule 5)- Every 5 conversational turns, or if an implementation path fails (bugs/regressions), you MUST halt and perform a context flush.- Explicitly ask the user to clear the chat history (CMD/CTRL + L or new Composer session).- Before resetting, summarize the current code delta and choices in a 5-line markdown bullet point, so the user can feed it into the fresh session alongside the updated full files.
- So those rules always work, we need to call the anchor at the beginning of any instruction/agent/chat like "#section2:lines 45-80" so the AI knows where to focus on. all right.
Have any of you guys already tried to implement a rule where AI identifies for itself (in a first msg for exemple) the files and the lines it needs to work on? I honnestly doubt this kind of last stuff would work, but who knows! It would be game changer to me lol
That's it !
Cheers & Thanks