dotclaude

Decompose a file that has grown over the size ceiling into focused sibling files — extracting at cognitive seams (hooks, sub-components, pure helpers, types), in the right order, behind a user-approved plan, with post-extraction verification. Invoke when a file-size check warns or a file has become too large to navigate. Refuses on auto-generated, snapshot, and fixture files.

The authority order across a project's three knowledge layers — guidance (.claude/), code (truth), and docs/ (reflection + transient intent) — with code as the tiebreak when a doc and the code disagree. Load when reading or trusting a doc to make a code decision, when authoring rules or skills that reference docs, or whenever a load-bearing fact comes from a description rather than from the code itself.

Core operating discipline for any non-trivial task — understand before building, reason to the right solution, execute completely and verified, stay lean, and avoid the parallel-path and subagent-dispatch traps. Load this on any substantive coding, design, planning, debugging, refactor, or review work (skip only for pure mechanics like a typo or an explicit one-liner). The universal "how to work" base, consumed as-is.

Set up worktree-per-feature discipline for a project where several AI sessions (or humans + agents) work concurrently and collide in one checkout. Writes a thin `worktree:` config block that the plugin's CONSUMED main-checkout hook reads (no rendered per-project hook copy), plus a project-specific worktree lifecycle skill — calibrated by interview (which repo, what's exempt, how a fresh worktree gets configured) and PROVEN by a live smoke worktree before handoff. Invoke /dotclaude:worktree when parallel sessions are real or planned.