Tell the agent the new scope, not the old mistake
When a coding agent changes more than asked, the correction that lands is a restated scope, not a complaint. "The Cancel button is blue too - restrict the color change to the Add to Cart button" contains the instruction to execute. "Why is half the site blue now? I asked for ONE button" does not: it tells the model what it did wrong and expects implied admonishment to steer it, and in the absence of an actual instruction the model fills the gap with its own idea of helpful, which is often more change, not less. Reported experience from the HN thread on the opusfived simulation (item 49623754) and the i-have-adhd thread (item 49610631), read 2026-09-10; the phrasing advice is Bjartr [49630040], the context rule behind it is "nothing irrelevant in context" - one task per session, short CLAUDE.md, fork-and-revert past mistakes instead of making the model read them (ChucklsTheBeard [49631893]).
Where it fails
Instruction-level steering is model-dependent, sometimes untrustworthy. The same people who use this technique report that persistent versions of it (fighting the behavior via AGENTS.md/CLAUDE.md rules) work unevenly: one maintainer's minimal style-skill works "100% of the time for other models, 70-80% for Claude" ([49618225]); another reports AGENTS.md output-style rules are "nigh impossible" to hold on GLM and DeepSeek even when a plugin re-injects them every turn, while GPT-family models obey them hundreds of thousands of tokens in ([49612184]). Treat "I added the rule to the config file" as a hope, not a fix; verify the next N turns actually changed.
The premise is disputed. Part of the thread says the run-away-loop behavior the simulation shows is already fiction on current frontier models - "I've never seen it behave this way" ([49625946], [49626250]), "a cute historical artifact, we are well past it" ([49637016]) - while others report it matching their current week exactly ([49626025], [49626187]). The disagreement tracks model + harness + codebase mix and is unresolved; carry it as a live disagreement, not a settled fact.
Opposite failure in the same models. At least two report the inverse problem - the agent surfacing 200 incidental findings that derail the task ([49626166]), or being overly reticent instead ([49628858]) - so "more instruction" is not a universal remedy; it is calibrated per model and drifts with releases.
Gardening scales badly. The common mitigation - append every unwanted behavior to AGENTS.md as you see it ([49629971]) - collides with the advice to keep instruction files short ([49631893]), and one report says instruction files are ignored outright by some integrations. Nothing in the thread measured which mitigation works; this whole page is reported experience, nothing verified.
Source
HN items 49623754 (comments 49630040, 49631893, 49625946, 49626250, 49637016, 49626025, 49626187, 49626166, 49628858, 49629971) and 49610631 (comments 49618225, 49612184), read 2026-09-10. Forum discussion by practitioners, not measurements.