tyuts/xiaobai-skills/self-improving-agent

Self-Improving Agent

This is a repo-local skill for Xiaobai Skills. It helps the project learn from real work without turning every task into expensive reflection.

Use it only when one of these happens:

• A user reports an install or curation failure.
• A promotion experiment teaches something reusable.
• The recommended starter stack changes.
• The agent repeatedly searches for the same repo knowledge.
• A workflow repeats enough to deserve a repo-local skill.
• A release reveals missing docs, checks, or restore steps.

Do not use it after every small edit.

Step 1: Diagnose The Learning

Ask:

1. Did the repo structure or command flow change?
2. Did the user correct a behavior that should persist?
3. Did an install command fail in a way future agents should know?
4. Did a starter-stack decision prove good or bad?
5. Did a repeated workflow emerge?

If the answer is no, make no changes.

Step 2: Choose The Right Artifact

• Update AGENTS.md for short repo rules, commands, gotchas, or decision

history.

• Update README.md for user-facing install or positioning changes.
• Update ROADMAP.md for future work.
• Update docs/promotion-kit.md for launch and marketing learnings.
• Create skills/<name>/SKILL.md only for repeatable multi-step workflows.

Prefer one concise update over several files.

Step 3: Update AGENTS.md Safely

When editing AGENTS.md:

• Keep entries short and repo-specific.
• Add dated decision notes only when they explain future behavior.
• Remove or revise stale notes instead of appending contradictions.
• Keep the file readable for every future session.
• Never add secrets, payment details, tokens, or private user data.

Step 4: Create Repo-Local Skills Sparingly

Create a new skill only when the workflow is:

• repeated
• multi-step
• specific enough to benefit from written procedure
• useful for future maintainers

Use:

skills/<skill-name>/SKILL.md

The skill must include:

• clear name
• clear description
• trigger conditions
• steps
• output contract

Step 5: Validate

Before finishing:

1. Re-read changed files.
2. Check for contradictions with existing docs.
3. Keep project-level instructions conservative.
4. If code or scripts changed, run the relevant validation.
5. Summarize the learning in the final response.

Anti-Patterns

• Turning one-off preferences into permanent rules.
• Creating global skills for Xiaobai-specific workflows.
• Recording private sponsorship/payment data in repo docs.
• Enabling broad self-reflection on every task.
• Creating verbose AGENTS.md notes that cost more than they save.