self-evolving/repo/distill-rubrics
Claude Code and agent skill by self-evolving/repo
- community
Checking that the SKILL.md file is still in the repository.
Distill durable user or team rubric preferences from local agent execution histories, repository discussions, and existing agent/rubrics content. Use when asked to mine prior agent work, compare candidate preferences with current rubrics, and open a PR that adds or updates rubric YAML on the agent/rubrics branch.
Install
npx vskill@latest install self-evolving/repo/distill-rubricsSKILL.md
Distill Rubrics
Use this skill to turn prior agent execution evidence into durable user/team rubrics. The output is a pull request whose base branch is agent/rubrics.
Workflow
- Determine the repository slug and rubrics ref.
- Default rubrics ref:
agent/rubrics. - Honor
AGENT_RUBRICS_REFwhen it is set. - Confirm the branch exists with
git ls-remote --heads origin "${AGENT_RUBRICS_REF:-agent/rubrics}". - If the branch does not exist, stop and report that rubric initialization must run first.
- Search local agent execution histories for this project.
- Prefer project-scoped evidence. Filter by repo slug, repo path, PR number, issue number, or branch name before trusting a hit.
- Search likely local history roots when they exist:
~/.codex/sessions/**/*.jsonl~/.acpx/sessions/*.json~/.acpx/sessions/*.stream.ndjson~/.claude/projects/**/*.jsonl- Also inspect mounted memory or history artifacts when present, especially
${MEMORY_DIR}/github/*.json. - Useful search terms include
rubric,preference,prefer,always,never,must,should,avoid,review,style, andworkflow.
- Convert history hits into durable evidence.
- Treat local histories as discovery, not final proof.
- For each promising hit, find the corresponding GitHub issue, PR, review, or comment with
gh. - Prefer stable URLs in
examples[].source, such as PR review comments or issue comments. - Do not learn rubrics from agent-authored summaries unless a trusted human explicitly endorsed the preference.
- Read the existing rubrics before editing.
- Create a temporary worktree or checkout rooted at the rubrics ref.
- Read every active rubric under
rubrics/. - Validate with
.agent/dist/cli/rubrics/validate.js --dir <rubrics-worktree>when the project runtime is built. - Prefer updating an existing rubric over creating a near-duplicate.
- Decide whether a rubric change is warranted.
- Add or update a rubric only when trusted evidence shows a durable preference future agents should follow or be evaluated against.
- Trust owner, admin, maintain, member, or collaborator feedback more than drive-by comments.
- Skip one-off remarks, speculative ideas, repository facts, and preferences already covered by an active rubric.
- Use
status: draftwhen the preference is plausible but not yet strongly established.
- Branch from the rubrics ref and edit only rubric files.
- Create a branch such as
rubrics/distill-<short-topic>fromorigin/agent/rubrics. - Store one rubric per YAML file under
rubrics/<area>/. - Keep directory names organizational; schema fields are the source of truth.
- Do not edit repository source files in the rubrics worktree.
- Validate, commit, push, and open a PR.
- Run rubric validation before committing.
- Commit only the rubric YAML or rubrics README changes needed for the distilled preference.
- Push the branch and open a PR with
gh pr create --base "${AGENT_RUBRICS_REF:-agent/rubrics}" --head <branch>. - Do not merge the PR unless the user explicitly asks.
Rubric Schema
Allowed domains are coding_style, coding_workflow, communication, and review_quality. Allowed routes are answer, implement, create-action, fix-pr, review, skill, rubrics-review, rubrics-initialization, and rubrics-update. Allowed severities are must, should, and consider; reserve must for clear repeated requirements.
Final Response
Report:
- rubrics branch checked
- local history sources searched
- candidate preferences considered
- rubric files created or updated
- validation result
- PR URL, or
no rubric changeswhen no durable preference was found
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryChecking.skills/distill-rubrics/SKILL.md
Popularity
- Repository stars
- 47
- Repository forks
- 4
- 7-day trend score
- 19
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/self-evolving/repo/distill-rubrics)