| name | repo-self-improvement |
| description | Use when repeated project mistakes, review feedback, failed commands, onboarding friction, or team workflow lessons should become reusable AI development behavior for this repo. |
Repo Self Improvement
Purpose
Turn repeated development lessons into reviewed, reusable project skills without letting agents rewrite shared behavior blindly.
When to Use
Use this skill after tasks that reveal:
- A repeated setup, build, test, or runtime failure.
- A reviewer correction that should become default behavior.
- A project convention missing from
AGENTS.md, .cursorrules, docs, or .agents/skills/.
- Onboarding friction that caused avoidable confusion.
- A tool-specific lesson from Hermes Agent, Codex, Cursor, VS Code, OpenCode, or another IDE.
Do not use it for one-off mistakes, personal preferences, or rules that can be enforced mechanically by lint, tests, or CI.
Core Rule
Agents may propose repo behavior improvements, but they must not silently overwrite shared team behavior. Proposals go through .agents/skill-proposals/ first unless the user explicitly asks to edit an existing skill.
Workflow
-
Identify the lesson.
- Name the concrete failure, correction, or repeated friction.
- Include evidence: command, error, file path, PR comment, or user correction.
-
Classify the destination.
AGENTS.md: strict repo-wide operating rule..cursorrules: guard rail for all AI assistants..agents/skills/<name>/SKILL.md: reusable workflow or judgment pattern.docs/: human-facing setup, architecture, or contribution guidance.- CI/lint/test: deterministic enforcement.
-
Draft a proposal under .agents/skill-proposals/.
- Use filename format
YYYY-MM-DD-short-topic.md.
- Keep it short and reviewable.
- Include the proposed target file and exact behavior change.
-
Validate before promotion.
- Check for conflict with
AGENTS.md, .cursorrules, and existing matching skills.
- Prefer updating an existing skill over creating a duplicate.
- Run relevant local checks if the proposal changes executable behavior.
-
Promote only after approval.
- Move accepted behavior into the target skill or doc.
- Keep the final skill concise.
- Remove stale or contradictory text when replacing old guidance.
Proposal Template
# Skill Proposal: <topic>
Date: YYYY-MM-DD
Target: <AGENTS.md | .cursorrules | .agents/skills/name/SKILL.md | docs/path.md | CI>
## Trigger
What happened, with concrete evidence.
## Proposed Behavior
What future agents or developers should do.
## Why It Belongs Here
Why this should be remembered instead of handled as a one-off.
## Validation
How to confirm the proposal is correct and non-conflicting.
Quality Bar
- Keep proposals small enough to review in one pass.
- Use repo-specific evidence, not generic best practices.
- Preserve tool neutrality unless a tool-specific behavior matters.
- Mention Hermes Agent when it improves learning or feedback loops, but do not make it mandatory for contribution.
- Favor commands and files the team already uses.
