← ClaudeAtlas

comment-writinglisted

Decides whether a comment is needed, then writes it short — class, function and inline alike — or reviews existing ones. No external docs, no bug hunting.
biancoenrico/claude-skills · ★ 0 · Code & Development · score 68
Install: claude install-skill biancoenrico/claude-skills
# comment-writing Writes the comments of a piece of code. The real work is **deciding which ones exist**, not writing the others. Generic — no conventions of its own, reads the project's in Step 1. `/dev-loop:code-revision` delegates its comment phase here for one copy of the criteria; in writing mode it also runs alone while the code is written. **Runs in a fork**, so its instructions and whatever it loads stay out of the caller's context and it never sees the conversation: the scope, and any *why* known only there, arrive in the arguments. Inside the loop there is no such argument — the whys it may use are what the code, the batch file and the plan already carry. Every question is handed back: `status: question` per `${CLAUDE_PLUGIN_ROOT}/references/agent-return.md`, report in `findings`, resume needs in `state`. **Below the small batch threshold** in `${CLAUDE_PLUGIN_ROOT}/references/limits.md` — measured with `sh ${CLAUDE_PLUGIN_ROOT}/scripts/devloop-diffsize --with-tests` over the scope — the caller follows this file **inline**: a fork and a reviewer cost more than a handful of comments. Inline, this overrides the fork rules above: whys may also come from the conversation, a question goes into the caller's own return, or to the user via `${CLAUDE_PLUGIN_ROOT}/references/asking.md`, and the point reached — candidates, verdicts, step — feeds the caller's `state`. ## The failure it exists to avoid A missing comment gets noticed and added; what doesn't is the opposite