kanlac/meta-learning

Meta Learning

Use this skill when "learning" means changing future agent behavior, not merely summarizing a source.

Goal

The best learning makes the owning skill clearer, lower-noise, and easier to execute. A good update changes the decision structure an agent will follow next time: a routing table, state machine, checklist, trigger rule, negative example, or sharper workflow. It is not a scrapbook of the source material.

Workflow

1. Name the future behavior: what should the next agent do differently?
2. Find the owning skill: update an existing skill when the lesson fits its domain. Create a new skill only when the behavior is reusable across domains or no owner exists.
3. Diagnose the weak decision point: ask what made the old behavior ambiguous. Typical fixes are "where to route this request", "when to escalate tools", "what not to create", or "what must be verified".
4. Refactor, do not append: rewrite the relevant section so the skill reads as a better operating guide. Merge sections that describe the same decision; remove or compress older text that the new structure replaces.
5. Keep the surface minimal: a few sentences belong in SKILL.md; a new reference file is justified only when the detail is optional, long, variant-specific, or too bulky for the main skill.
6. Verify the learning: check for dangling links, private facts, hardcoded local paths, line-count drift, and whether plugin README/manifests need updates.

Patterns

• Low-information source: extract one principle or 2-3 examples; do not mirror the source outline.
• Tool/source choice lesson: prefer a routing table over prose because it makes execution explicit.
• User correction: update the trigger or anti-pattern that allowed the mistake, then simplify nearby text.
• Site-specific mechanics: keep login/browser details in the operational skill; only split to references when the steps are too long for the main document.
• Terminology: use domain-native words. Do not invent vague jargon, especially mixed English labels, unless the ecosystem already uses them.

Anti-Patterns

• Keeping two sections that describe one decision because they came from different notes.
• Creating a new reference file because a source had headings.
• Copying a report's taxonomy when a single rule would change behavior.
• Adding project-private names, paths, cases, clients, or environment assumptions to a public skill.