explain

Teach one gnarly topic (a bug family, a design constraint, a system) until you genuinely understand it: author a verified, layered explainer doc through an isolated investigation, then walk it with yo

SKILL.md

GitHub

Two modes. If $ARGUMENTS names an existing explainer doc, run it. If it names a topic, author one first, then offer to run it.

$ARGUMENTS

Author

Delegate the investigation to sy:explain-author — an isolated context, so its file reads and repro attempts never enter this session. Seed the prompt with every anchor already in hand: paths, symbols, the commit, the plan/gate/CI finding that made this worth explaining. It writes the doc itself to <topic>_explainer.md in the topic's resolved scratch directory (scratch_dir(<topic-slug>)) and returns only a compact brief — the resolved doc path, layer count, the one surprising rule, and how many claims are verified versus inferred.

Present that brief as a Status update, then close with one optional suggestion (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/user-interaction.md): walk it here now, or run claude "/sy:explain <doc path>" in a new session for a clean slate before something like a merge. Never block on it — proceed either way.

Run

references/running.md owns the walk: one layer per turn, an AskUserQuestion checkpoint after each, live repros the moment something is challenged, a decision (not a fact) at the end.

No tracker read or write, at any point — this is about your understanding, not the project record.

Every subagent dispatch resolves its model from config and passes it as the Agent invocation's actual model override, per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/model-dispatch.md; a nested dispatch inherits nothing and must resolve again.