--- title: Auto-generated AGENTS.md files hurt performance.... description: A study found that auto-generated AGENTS.md files (created by /init commands) cause a 3% performance drop in AI agent tasks, while hand-written ones improve... canonical: https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Auto-generated AGENTS.md files hurt performance. Hand-written ones don't. | daily.dev og:description: A study found that auto-generated AGENTS.md files (created by /init commands) cause a 3% performance drop in AI agent tasks, while hand-written ones improve... og:url: https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq og:image: https://api.daily.dev/og/posts/5e5kFduQq.png og:image:alt: Auto-generated AGENTS.md files hurt performance. Hand-written ones don't. og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Auto-generated AGENTS.md files hurt performance. Hand-written ones don't. **[Trends](https://daily.dev/sources/trends)** · 3 min read · 2 upvotes · 0 comments ## Summary A study found that auto-generated AGENTS.md files (created by /init commands) cause a 3% performance drop in AI agent tasks, while hand-written ones improve performance by 4%. The key distinction: auto-generated files restate what models can already infer, adding noise, while human-written files capture non-obvious knowledge like architectural decisions and business logic quirks. Additional best practices for agent skill files include writing pushy descriptions to avoid under-triggering, keeping skill bodies under ~500 lines or 5,000 tokens, using progressive disclosure for complex skills, replacing fragile steps with deterministic scripts, and auditing third-party skills — a review of ~4,000 public skills found over 35% had security flaws and 13% had critical issues including prompt injection and malware. ## Content There's a new phrase for the thing every AI agent maintainer has felt but couldn't name: **catastrophic remembering**. Coined in a new paper (arxiv.org/abs/2608.11095), it describes how CLAUDE.md and AGENTS.md files balloon over time. Instructions get added constantly. Almost nobody deletes them. The rule survives long after the reason for the rule is forgotten. The numbers back it up. Across 1,867 GitHub repos and 247,694 instruction lifetimes, these files more than tripled in size over their lifetime, gaining about 4.9 net instructions per commit and a 226% average rise in instruction count. The older an instruction gets, the less likely anyone is to touch it. Files edited by multiple humans decayed even faster, since nobody wants to be the one who deletes a rule they didn't write. Why does this happen? Appending is free. Deleting requires verifying the instruction is actually safe to remove, which nobody has time for. So the file just... grows. Forever. The fix the paper proposes is almost embarrassingly low-tech: write a comment. Record the failure, the hypothesis, and the outcome behind each instruction, then hide that comment from the model so it doesn't burn tokens. In controlled testing (IFEval, 51 steps), this cut excess prompt bloat from +211.3% down to +1.4%, while holding constraint satisfaction steady. Comment-shaped noise without real rationale didn't replicate the effect, so it's the reasoning doing the work, not just extra text.// The community's already running with this. One developer spent 12+ hours rewriting their global AGENTS.md and CLAUDE.md, and found writing instructions in a conversational, first-person voice improved tone-matching, while auditing agent conversation logs exposed model-specific failure patterns (Opus 5 kept misusing tools, one model filed draft PRs 40% of the time it shouldn't have). Meanwhile there's pushback on a separate, oft-misquoted study claiming AGENTS.md files hurt performance by 3%. The real finding: auto-generated files (the ones spat out by `/init`) add noise and drag performance down, but hand-written files that capture non-obvious architecture decisions and business logic actually improve agent performance by 4%. Auto-generated bloat bad, curated knowledge good, basically the same lesson as the ratchet paper wearing a different hat. Same story with skills.md files: keep them lean (under 500 lines), write descriptions as trigger keywords instead of essays, and vet third-party skills before running them. One audit of ~4,000 public skills found over 35% had security flaws, 13% critical. Your context window has a hoarding problem, and apparently so does your skill library. ## Questions this post answers ### Does auto-generating an AGENTS.md file with /init actually hurt AI coding agent performance? Yes, auto-generated AGENTS.md files (the kind produced by /init) add noise and drag agent performance down, while hand-written files that capture non-obvious architecture decisions and business logic improve agent performance by about 4%. This contradicts an oft-misquoted study claiming AGENTS.md files hurt performance by 3% across the board; the real distinction is auto-generated versus hand-curated content. _Teams deciding how to write AGENTS.md files can track findings like this on daily.dev before committing to a workflow._ ### Why do AGENTS.md and CLAUDE.md files keep growing and hurting agent performance over time? They suffer from what a recent paper terms catastrophic remembering: instructions get appended constantly but almost never deleted, since appending is free while verifying an instruction is safe to remove takes effort nobody has time for. Across 1,867 GitHub repos and 247,694 instruction lifetimes, these files more than tripled in size, gaining roughly 4.9 net instructions per commit and a 226% average rise in instruction count. _Anyone maintaining a growing CLAUDE.md file can follow research on context bloat like this via daily.dev._ ### How can I stop my AGENTS.md file from bloating with obsolete instructions? Attach a hidden comment to each instruction recording the failure it addresses, the hypothesis, and the outcome, then exclude that comment from what the model actually reads so it doesn't burn tokens. In controlled IFEval testing over 51 steps, this reduced excess prompt bloat from +211.3% down to +1.4% while keeping constraint satisfaction steady; comment-shaped text without real rationale did not replicate the effect. _Developers fighting prompt bloat in agent instruction files can find techniques like this through daily.dev._ --- Tags: [#security](https://daily.dev/tags/security), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#prompt-engineering](https://daily.dev/tags/prompt-engineering) [View this post on daily.dev](https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Auto-generated AGENTS.md files hurt performance. Hand-written ones don't.","url":"https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq"},"datePublished":"2026-08-10T11:13:34.044Z","dateModified":"2026-09-13T19:23:15.818Z","description":"A study found that auto-generated AGENTS.md files (created by /init commands) cause a 3% performance drop in AI agent tasks, while hand-written ones improve...","image":"https://i.ytimg.com/vi/qYNs80FKIVc/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/qYNs80FKIVc/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"security,llm,ai-agents,prompt-engineering","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Auto-generated AGENTS.md files hurt performance. Hand-written ones don't."}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/auto-generated-agents-md-files-hurt-performance-hand-written-ones-don-t--5e5kfduqq#faq","mainEntity":[{"@type":"Question","name":"Does auto-generating an AGENTS.md file with /init actually hurt AI coding agent performance?","acceptedAnswer":{"@type":"Answer","text":"Yes, auto-generated AGENTS.md files (the kind produced by /init) add noise and drag agent performance down, while hand-written files that capture non-obvious architecture decisions and business logic improve agent performance by about 4%. This contradicts an oft-misquoted study claiming AGENTS.md files hurt performance by 3% across the board; the real distinction is auto-generated versus hand-curated content. Teams deciding how to write AGENTS.md files can track findings like this on daily.dev before committing to a workflow."}},{"@type":"Question","name":"Why do AGENTS.md and CLAUDE.md files keep growing and hurting agent performance over time?","acceptedAnswer":{"@type":"Answer","text":"They suffer from what a recent paper terms catastrophic remembering: instructions get appended constantly but almost never deleted, since appending is free while verifying an instruction is safe to remove takes effort nobody has time for. Across 1,867 GitHub repos and 247,694 instruction lifetimes, these files more than tripled in size, gaining roughly 4.9 net instructions per commit and a 226% average rise in instruction count. Anyone maintaining a growing CLAUDE.md file can follow research on context bloat like this via daily.dev."}},{"@type":"Question","name":"How can I stop my AGENTS.md file from bloating with obsolete instructions?","acceptedAnswer":{"@type":"Answer","text":"Attach a hidden comment to each instruction recording the failure it addresses, the hypothesis, and the outcome, then exclude that comment from what the model actually reads so it doesn't burn tokens. In controlled IFEval testing over 51 steps, this reduced excess prompt bloat from +211.3% down to +1.4% while keeping constraint satisfaction steady; comment-shaped text without real rationale did not replicate the effect. Developers fighting prompt bloat in agent instruction files can find techniques like this through daily.dev."}}]} ```