review-rules
Scannednpx machina-cli add skill SenZhangAI/your-taste/review-rules --openclawReview rule proposals that your-taste has extracted from your sessions.
Important: Always respond in the user's language (infer from their recent messages in this conversation).
Steps
- Get proposals data:
node "${CLAUDE_PLUGIN_ROOT}/bin/cli.js" review-data
-
Parse the
proposalsarray from the output. If empty, tell the user: "No rule proposals yet. Keep using Claude Code and proposals will surface when patterns emerge." -
For each proposal, present it to the user with:
- The rule text
- The evidence
- The source (session file or taste:init) Ask: "Accept, Edit, or Dismiss?"
-
Collect all decisions into this JSON format:
{
"accepted": ["rule text 1", "rule text 2"],
"edited": [{"original": "old text", "revised": "user's edited text"}],
"dismissed": ["rule text 3"]
}
- Apply decisions:
echo '<the JSON above>' | node "${CLAUDE_PLUGIN_ROOT}/bin/cli.js" review-apply
- Report results: how many rules were approved and added to CLAUDE.md.
Source
git clone https://github.com/SenZhangAI/your-taste/blob/main/skills/review-rules/SKILL.mdView on GitHub Overview
This skill helps you review rule proposals extracted from your-taste sessions, validate them with evidence, and decide to Accept, Edit, or Dismiss. It guides you through presenting each proposal with its rule text, evidence, and source, and then applying the chosen decisions to CLAUDE.md.
How This Skill Works
Fetch proposals with the CLI (node "${CLAUDE_PLUGIN_ROOT}/bin/cli.js" review-data) and parse the proposals array. For each proposal, present the rule text, evidence, and source and ask the user to Accept, Edit, or Dismiss. Finally, submit the collected decisions as JSON to the review-apply command to update CLAUDE.md.
When to Use It
- When a new rule proposal surfaces from your sessions and needs validation.
- When you want to prune CLAUDE.md by removing obsolete or conflicting rules.
- When multiple proposals exist and you need consistent wording before approval.
- When collaborating with teammates to decide on rules and maintain traceability.
- When you require provenance and a record of source and edits for each rule.
Quick Start
- Step 1: Retrieve proposals with node "${CLAUDE_PLUGIN_ROOT}/bin/cli.js" review-data
- Step 2: Review each proposal and choose Accept, Edit, or Dismiss; collect into the accepted, edited, and dismissed arrays
- Step 3: Apply decisions by piping the JSON to the CLI: echo '<json>' | node "${CLAUDE_PLUGIN_ROOT}/bin/cli.js" review-apply
Best Practices
- Verify that each rule text matches the intended behavior described by the evidence.
- Limit edits to preserve original intent; note changes in the 'edited' entries.
- Keep sources and evidence visible to maintain auditability.
- Use concise, unambiguous language and consistent formatting.
- After applying decisions, review CLAUDE.md to confirm the changes are correct.
Example Use Cases
- Rule: 'Do not disclose secrets in any output'; Evidence: repeated prompts across sessions; Source: session-23.json
- Rule: 'Require explicit user confirmation before destructive actions'; Evidence: best practice note in taste:init; Source: taste:init
- Rule: 'Flag proposals that widen access permissions'; Evidence: several sessions flagged for review; Source: session-19.json
- Rule: 'Use clear and consistent rule wording'; Evidence: inconsistent wording observed across proposals; Source: session-25.json
- Rule: 'Route ambiguous proposals to human review'; Evidence: user feedback suggesting escalation; Source: session-42.json
