Tips Tip #1 and Tip #2 covered setting up copilot-instructions.md (or AGENTS.md) and growing it over time. But can it grow too much?
Some early research suggests yes. A couple of recent studies (ICSE JAWs 2026, ETH Zurich) found that auto-generated instruction files — the kind produced by /init commands that scan your codebase and describe what they find — actually hurt performance slightly while increasing cost. Meanwhile, hand-written files with genuinely non-obvious info (“use uv not pip”, “the auth middleware is fragile — don’t refactor it”) seemed to help.
It’s early days and these are small studies, so take the numbers with a grain of salt. But the intuition rings true: repeating what’s already in the code is noise. Telling the AI what it can’t discover on its own is signal.
A useful filter to try: before adding a rule, ask yourself — would the AI figure this out anyway from the code and tests? If yes, maybe skip it.
💡 Try this: Review your instruction file (or a teammate’s). Is anything in there that the AI would learn just by reading the code? Consider trimming it.
🔗 Addy Osmani’s write-up with links to both studies
Tried pruning your instruction file(s)?
🟢 Regularly — it’s part of our workflow
🟡 Yes, once or twice
🔴 Not removed much yet, no
⚪ Instruction file … ? 😉
Have you removed something that you then discovered had to be in there?