Capture Learning

Create one source-grounded solution document that a later engineer can find and use. The next consumer is someone facing the same problem. Done means the document is written, validated, and does not state an unverified claim as fact.

Scope

Capture one learning per run. It must describe a solved, verified, and non-obvious problem. If the evidence does not meet that bar, do not write a weak document. In mode:headless, return Documentation skipped with the reason. In an interactive run, ask only when the solved problem itself is unclear.

This skill writes only the solution document. When an existing CONCEPTS.md has a qualifying missing term, update that entry too. Do not create a glossary, edit project instructions, search unrelated memory, or turn this run into a documentation audit.

Workflow

1. Read the relevant conversation evidence, changed source, verification results, and nearby solution documents. Source and captured command results outrank session recollection.
2. Read references/schema.yaml and references/yaml-schema.md to classify the learning and choose its destination. Read references/resolution-template.md for the required body shape.
3. Check for an existing document about the same problem. Update it when the problem, cause, and solution materially match; otherwise create one new document. Keep the path stable when updating.
4. Ground every behavior claim in current source with a file:line reference. Describe merge state only when it is verified. Attribute or remove claims that cannot be verified.
5. Read references/concepts-vocabulary.md only when an existing glossary may need a qualifying term. Keep the glossary change within the learning's domain.
6. Run scripts/validate-frontmatter.py and scripts/validate-doc-claims.py from this skill's directory. Read references/grounding-validation.md to adjudicate flags. Fix, annotate, or confirm each flag; never silently ignore one. If a validator cannot run, perform its documented manual checks and report that fallback.