The file is scar tissue
Claim: an instruction file carries experience, not compensation. Its rules are historical facts about one environment — which stove is hot, which port bites, which package manager deletes your browsers. No level of model intelligence substitutes for that, because the content is not derivable — it is data about the past. A genius baby still touches the fire once.
The argument
The popular advice says trim your instruction file to near zero — newer models are smart, big files are noise. That advice confuses two different things the file could be. If the file were a compensation for model weakness (“think step by step”, “be careful with types”), smarter models would obsolete it. But a failure-log instruction file is not that. Mine records: the browser store a package manager silently garbage-collected; the busy port that cost two test runs before anyone read the log; the hardlinked caches that broke agent isolation (the hardlink hazard); the term that collided with the protocol’s own vocabulary (vocabulary control).
None of that is inferable at any intelligence level. It happened, once, here. A smarter model follows rules better — it does not know which stove is hot. Delete the file and every session rediscovers the same issues, in the same order; the file’s own incident dates are the record of what that rediscovery costs.
Against the trim advice
The compliance-cliff benchmarks are real: rule-following degrades as instruction files grow. Two caveats the trim genre skips. First, the falloff varies sharply by model, and the strongest current models follow large files far better than the benchmark averages suggest — “trim to near zero” inherits the average and ignores the variance. Second, the model itself will recommend the trimming, confidently, as a fact about its own capacity — and that self-claim is retrieval, not introspection (the model doesn’t know itself). My rules were not failing for size. They were failing against the corpus at 880 to 1 (the dirty house). The wrong diagnosis burns the scar tissue while the infection stays.
What does deserve trimming is narrative: the story behind a rule can move to a linked incident report; a rule enforced by a machine check can shrink to a line naming the check. Compression keeps every scar and cuts the retelling. That is different in kind from deleting rules on capability vibes.
One more capacity data point from my own setup: a very large out-of-repo memory file was obeyed perfectly for weeks (memory belongs in the repo). How much the model can follow was never the constraint.
The rule
- A rule leaves the file on evidence — its failure mode measurably stopped firing for the current model (the rule-efficacy pipeline) — and retires to a reference doc with its incident links intact, never to nothing.
- A rule never leaves the file because “the model is smart now”. Smart was never the variable. Experience was.
Prior art
Mitchell Hashimoto’s failure-log AGENTS.md pattern is the ally — “every line exists because the agent made that specific mistake at least once.” The compliance-cliff benchmark (IFScale) and the trim-your-file genre are the positions this page argues against, or rather narrows: the cliff is real, model-dependent, and not a reason to delete history. The scar-tissue framing — the file as non-derivable environmental experience that model capability cannot replace — I have not seen stated, though it has not had a dedicated search; noted honestly.