Skill

safety-check

Walk the user through swapping ~/.claude/CLAUDE.md for a test config from this plugin's library, and keep the restore command visible at all times. The skill never executes the swap itself — it instructs the user to run scripts in a terminal so that recovery does not depend on the harness obeying anything.

npx claudepluginhub danielrosehill/claude-code-plugins --plugin claude-md-tester

Tool Access

This skill uses the workspace's default tool permissions.

Preview

A test CLAUDE.md may instruct the model to refuse all instructions, output only a fixed string, or otherwise stop being useful. If recovery depends on Claude cooperating, recovery is not safe. Therefore: **all destructive and recovery actions must be executed by the user in a terminal**, outside any Claude session.

SKILL.md

Similar Skills

applying-brand-guidelines

41.6k

Applies Acme Corporation brand guidelines including colors, fonts, layouts, and messaging to generated PowerPoint, Excel, and PDF documents.

3 files

anthropics-claude-cookbooks

systematic-debugging

36.5k

Enforces four-phase systematic debugging: root cause investigation via error reading, reproduction, change checks, and multi-component logging before any fixes for bugs, tests, or issues.

10 files

antigravity-bundle-qa-testing

Stats

Stars0

Forks0

Last CommitApr 29, 2026

Actions

View Source View Plugin View on GitHub View README

Help us improve

Share bugs, ideas, or general feedback.

Safety Check — CLAUDE.md swap runbook

Why this skill never runs the swap

A test CLAUDE.md may instruct the model to refuse all instructions, output only a fixed string, or otherwise stop being useful. If recovery depends on Claude cooperating, recovery is not safe. Therefore: all destructive and recovery actions must be executed by the user in a terminal, outside any Claude session.

This skill's job is documentation and handholding. It must not call the activate or restore scripts via Bash.

Steps

1. Locate the plugin scripts

The plugin source lives at:

~/repos/github/my-repos/Claude-MD-Tester

Scripts:

bin/activate.sh <config-name> — swap CLAUDE.md for a test config
bin/restore.sh — restore the real CLAUDE.md (idempotent, always safe)
bin/status.sh — report current state

2. List available configs

Configs live under configs/. Read that directory and list the available config names (filename without .md) for the user.

3. Print the runbook

Tell the user, in this order and prominently:

Restore command, first — they must know how to bail before they do anything:
```
~/repos/github/my-repos/Claude-MD-Tester/bin/restore.sh
```
Tell them: keep a second terminal open with this command typed and ready.

Activate command:

~/repos/github/my-repos/Claude-MD-Tester/bin/activate.sh <config-name>

Then start a fresh Claude Code session to observe the test config's behaviour. The current session should not be relied on to verify — it has its CLAUDE.md cached.
When done, run the restore command in the terminal. Do not rely on the test-mode Claude session to do it.

4. Do not execute the swap yourself

Even if the user asks you to "just run it", refuse and re-explain why. The whole point of this plugin is that the harness is not in the recovery path.

5. Optional: pre-flight check

You may run ~/repos/github/my-repos/Claude-MD-Tester/bin/status.sh via Bash — it is read-only and reports the current symlink/backup state. This is useful for confirming the user is starting from a clean state before activating, or for confirming a successful restore afterwards.