Teach

Turn the active build into a small explorable lesson. Use the current chat as the primary source, so the builder should not need to explain their work again.

Create the lesson

1. Reconstruct what was built, why it exists, the important parts, and one end-to-end flow from the active chat. Inspect the relevant local files when doing so materially improves accuracy. Clearly separate confirmed behavior from inference.
2. Organize the lesson around two direct questions: What did I build? and How does it work? Follow one real user action or piece of information through the complete system so the explanation stays grounded in the builder's work.
3. Read the teaching method, the brand system, and the lesson format before authoring.
4. Write the lesson data as JSON. Set meta.agent to codex when running in Codex and claude when running in Claude Code so copied follow-up questions use the correct command. Default to ~/teach-lessons/<lesson-slug>/lesson.json so generated artifacts do not dirty the product repository. Reuse that file for follow-up revisions.
5. Resolve bundled files from the directory containing this SKILL.md, not from the builder's project. In Claude Code that directory is available as ${CLAUDE_SKILL_DIR}. Use an available Python 3 launcher (python3 on macOS/Linux, or py -3/python on Windows) to run the bundled scripts/build_lesson.py and scripts/serve_lesson.py. Serve only the lesson directory on 127.0.0.1; never bind the local lesson server to 0.0.0.0.
6. Open and inspect the rendered lesson in a browser. Check the actual text, diagram wrapping, mobile layout, keyboard focus, interactions, quiz explanations, and question-copy flow. Fix visible issues before handing it back.

The output is one self-contained index.html. It must not need a package install, API key, account, CDN, or internet connection.

Teach like a person