Releases·Tool Release
Your CLAUDE.md can reference code that no longer exists. I made a small pre-commit check for it
A developer released check-docs, a 39-line sh pre-commit script that catches when CLAUDE.md or AGENTS.md instruction files reference files or components that no longer exist. Stale agent instructions make Claude appear to hallucinate when the real problem is outdated project context. The tool runs with no model calls, packages, or services.
1 engagement1 sourceFri, Sep 18, 2026, 07:59 PM
AI Overview
- ipaulsmith's 39-line sh script check-docs verifies CLAUDE.md/AGENTS.md referenced paths exist, deleted names removed, and instruction changes staged, with no model call
Related
Social·Community
Developers report CLAUDE.md context files go stale in weeks without maintenance
Multiple developers report that CLAUDE.md and similar context files for AI coding agents quickly become outdated, leading agents to act on obsolete rules. A small team of five found their file drifted after a month despite prompts to update it. Others note that ineffective files often contain vague instructions rather than actionable commands. The community is seeking better maintenance strategies.
Multiple developers report that CLAUDE.md and similar context files for AI coding agents quickly become outdated, leading agents to act on obsolete rules. A small team of five found their file drifted after a month despite prompts to update it. Others note that ineffective files often contain vague instructions rather than actionable commands. The community is seeking better maintenance strategies.
Updated Sun, Jul 26, 2026, 04:13 PM — Users share tips: actionable instructions (e.g., 'Never use npm') keep context files effective, while vague statements l
131110 engagement5 sourcesredditTue, Jul 21, 2026, 07:47 PM
Social·Community
What belongs in your CLAUDE.md because Claude got it wrong once?
A Reddit discussion asks developers which instructions they added to CLAUDE.md after Claude got something wrong, and whether every mistake deserves a permanent rule. The thread surfaces project-specific rules like checking current docs before config changes and scoping fixes to avoid touching unrelated code. It matters to practitioners as a collection of hard-won agent-guardrail practices and debate over rule maintenance.
A Reddit discussion asks developers which instructions they added to CLAUDE.md after Claude got something wrong, and whether every mistake deserves a permanent rule. The thread surfaces project-specific rules like checking current docs before config changes and scoping fixes to avoid touching unrelated code. It matters to practitioners as a collection of hard-won agent-guardrail practices and debate over rule maintenance.
53 engagement2 sourcesredditFri, Sep 18, 2026, 07:58 AM
Releases·Tool Release
agent-standard-oss: open-source tool scores and fixes AI agent config files in repos
A developer released agent-standard-oss, an open-source CLI tool that scores a repository's AI agent configuration files (AGENTS.md, CLAUDE.md) on a 6-point checklist and helps fix common issues like duplicate rules or missing git saves. The tool aims to improve code quality and agent output by ensuring repos have a single, up-to-date instruction file for AI assistants.
A developer released agent-standard-oss, an open-source CLI tool that scores a repository's AI agent configuration files (AGENTS.md, CLAUDE.md) on a 6-point checklist and helps fix common issues like duplicate rules or missing git saves. The tool aims to improve code quality and agent output by ensuring repos have a single, up-to-date instruction file for AI assistants.
8 engagement1 sourceredditThu, Jul 23, 2026, 02:10 AM
Social·Community
Anthropic is training 10,000 Claude engineers. My own agent audit found 85 stale instructions hiding under a strong model
After Anthropic announced Frontier Academy, a Reddit user reflected on what the program should teach and described auditing their own Claude-based agent one evening. The audit, prompted by feeling that Fable 5 made everything faster and safer, uncovered 85 stale or wrong references across 18 core docs plus five audit agents fanning out over the rest of the system. The user's core warning: a strong model routed around stale instructions so well that maintenance debt stayed invisible.
After Anthropic announced Frontier Academy, a Reddit user reflected on what the program should teach and described auditing their own Claude-based agent one evening. The audit, prompted by feeling that Fable 5 made everything faster and safer, uncovered 85 stale or wrong references across 18 core docs plus five audit agents fanning out over the rest of the system. The user's core warning: a strong model routed around stale instructions so well that maintenance debt stayed invisible.
7 engagement1 sourceredditSun, Oct 4, 2026, 12:42 PM
Social·Community
Investigation finds half of 6 AI coding tools write no plan to disk, causing agent forgetfulness
An investigation into six AI coding tools found that half of them write nothing to disk to persist their plans, which explains why agents forget what they were doing. The author argues the issue is a filesystem problem, not a context-window problem, and that checking where the plan is written predicts which tools suffer from the bug.
An investigation into six AI coding tools found that half of them write nothing to disk to persist their plans, which explains why agents forget what they were doing. The author argues the issue is a filesystem problem, not a context-window problem, and that checking where the plan is written predicts which tools suffer from the bug.
12033 engagement3 sourcesredditSun, Aug 16, 2026, 08:44 AM