learn

name	learn
description	Use when a user corrects, rejects, edits, or redirects an LLM/agent response and the correction should become a reusable reasoning strategy. Converts feedback into generalized learnings for ~/.agents/STEERING.md with linked RDF/Turtle evidence. USE FOR: user corrections, preference feedback, rejected agent behavior, reasoning strategy updates, steering file maintenance DO NOT USE FOR: storing task facts (use memory), ordinary skill authoring (use agent-skills), project instruction files unrelated to feedback (use agents-md)
license	MIT
metadata	{"displayName":"Learn","author":"Tyler-R-Kendrick","tags":["ai","feedback","steering","learning"]}
compatibility	claude, copilot, cursor

Learn

Use this skill to turn a user correction into a generalized steering rule. Keep STEERING.md as an index, and keep details in linked RDF/Turtle files under ~/.agents/.steering/.

This turns feedback loops into explicit, inspectable learnings. See correction-to-learning.md for the correction-to-learning model.

Route

• IF the user corrects an agent response, read correction-to-learning.md and zpd-teaching-loop.md.
• IF creating or updating steering artifacts, read steering-index-contract.md and rdf-contract.md.
• IF using the deterministic helper, read cli-usage.md and run node skills/ai/learn/scripts/dist/learn.js.
• ELSE answer normally and do not create steering entries.

Core Workflow

1. Identify the correction, the agent behavior it changes, and the broader class of future tasks it should affect.
2. Generalize the correction into a strategy or preference, not a one-off fact.
3. If memory is available, search for related prior corrections and add memory:// evidence links; keep facts in memory, not in steering.
4. Add evidence links to docs, instruction files, conversations, or other sources.
5. Use the CLI to write the RDF entry and regenerate STEERING.md.
6. Briefly teach the agent what reasoning move to practice next time.

Best Practices

• Always store durable learnings as generalized strategies or preferences, not isolated task facts.
• Keep STEERING.md as a conditional index and load linked RDF files only when the current task matches the entry.
• Require evidence links for every learning so agents can inspect provenance before relying on the rule.
• Use the zone of proximal development to explain the next reasoning move the correction teaches.
• Prefer deterministic CLI writes over manual steering edits to avoid broken links, duplicate IDs, or stale index rows.