Audit agent instruction files (CLAUDE.md, AGENTS.md, .claude/rules/) against the commits since any of them last changed, then propose updates each motivated by a commit
Installs into .claude/skills of the current project.
Are you the author of Agent Docs Drift?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/bengous-agent-docs-drift)
---
name: agent-docs-drift
description: Audit agent instruction files (CLAUDE.md, AGENTS.md, .claude/rules/) against the commits since any of them last changed, then propose updates each motivated by a commit
disable-model-invocation: true
license: MIT
allowed-tools:
- Bash("${CLAUDE_PLUGIN_ROOT}/scripts/agent-docs-drift.ts":*)
- Read
- Write
- Edit
- Agent
- AskUserQuestion
- Glob
- Grep
---
# Agent Docs Drift
Audits CLAUDE.md, AGENTS.md and `.claude/rules/` against the commits since any of them last changed. Never commits: the user commits.
1. Run the report:
```bash
"${CLAUDE_PLUGIN_ROOT}/scripts/agent-docs-drift.ts"
```
A non-zero exit prints the reason on stderr: relay it and stop. When no doc file is tracked, point the user to `/init`. An empty window: relay the report's result and stop.
2. Follow the report's protocol. In Claude Code:
- a subagent is an Agent tool call with `subagent_type: Explore`, briefed with the baseline, protocol step 3, the audit set and its share of commits;
- the approval question is an AskUserQuestion (apply all / cherry-pick / reject);
- general quality improvement of the docs belongs to `/context-audit`.
