Trigger Scenarios

检查注释质量审一下注释注释评审 optimize the comments

Install

npx skills add openkursar/hello-halo --skill comment-review -g -y

More Options

Non-standard path

npx skills add https://github.com/openkursar/hello-halo/tree/main/.claude/skills/comment-review -g -y

Use without installing

npx skills use openkursar/hello-halo@comment-review

指定 Agent (Claude Code)

npx skills add openkursar/hello-halo --skill comment-review -a claude-code -g -y

安装 repo 全部 skill

npx skills add openkursar/hello-halo --all -g -y

预览 repo 内 skill

npx skills add openkursar/hello-halo --list

SKILL.md

Frontmatter

{
    "name": "comment-review",
    "description": "Must be invoked when the user asks to review, audit, or improve code comments — e.g. \"check the comment quality\", \"审一下注释\", \"注释评审\", \"optimize the comments\". Reviews comments in the current changes against this project's comment policy and fixes the issues found."
}

Comment Review & Improvement Skill

Reviews the comments in the current changes against this project's comment policy (CLAUDE.md), reports violations, and — after confirmation — fixes them.

The governing principle (from CLAUDE.md): comments transfer context the reader cannot get from the code itself. Default to none. Keep each comment as short as the context allows. If removing the comment loses no context, it shouldn't exist.

Workflow

1. Collect the changed comments

Run git status and git diff to see uncommitted changes (or the range the user names, e.g. a branch/commit).
Focus only on comment lines that were added or modified. Ignore pure logic diffs and generated files (locale JSON, lockfiles, snapshots).
For large diffs, save the diff to a file and extract added comment lines, then read the surrounding code so each comment is judged in context — never judge a comment by its text alone.

2. Evaluate each comment

Read the code the comment sits on, then decide: keep / trim / delete / rewrite.

Keep (only if the code alone cannot convey it):

Why a non-obvious decision was made.
An invariant that must hold, or a trap not visible locally.
A security/trust boundary, a race-condition hazard, a non-destructive contract.
A link to an issue/RFC.

Flag for fix — these are violations of the policy:

#	Violation	Example	Action
1	Ceremonial / subjective wording	"The proud part of the design…", "elegant", "cleanly"	delete the editorializing
2	Planning labels / session IDs	`E1`, `F2`, `Phase X`, `Full-chain:`, `Helper:`	delete
3	Cross-function line references	"see api.apiKey above", "(see line 120)"	delete the pointer; state the rule once
4	Narrating removed/changed code	"new — replaces…", "legacy; was…", "the previous branch is gone"	state the present rule only
5	Paraphrasing adjacent code	`// user re-entered — keep it` above `if (val) return`	delete
6	Restating a condition	`// oldest first` is borderline; long restatements are not	trim or delete
7	Defensive notes aimed at reviewers	"Note: this is safe because…" addressed to a PR reader	delete or reframe as an invariant
8	Numbered cross-refs to a scheme defined elsewhere	"(invariant ②)"	name the property inline instead
9	Over-long blocks	multi-paragraph essays / worked examples where 2–3 lines suffice	halve it; keep the one-line why

Also check the basics from the commit policy:

Comments in English (open-source convention).
No internal/private info, no company/brand names.

3. Report before changing anything

Produce a concise report:

A density signal for the worst files (comment-lines / total-lines) to show over-documentation.
A prioritized list: clear violations (#1–#5) first, then long-block trims.
For each item: file:line, the rule it breaks, and the proposed shorter version.
Explicitly list what you will keep so valuable rationale is not stripped.

Do not delete-all. Stripping a security/why comment is worse than an over-long one.

4. Fix on confirmation

Per CLAUDE.md, code/comment edits require human consent. Wait for the user to approve (or to pick a subset) before editing.
Make comment-only edits — never alter logic. Prefer Edit on the exact block.
After editing, do a quick pass to confirm no {@link}/name references now dangle and no logic line was touched.

Notes

If there are no comment changes, say so — nothing to review.
Do not run git commit or git push; this skill only reviews and edits comments.
Read CLAUDE.md (and halo-dev) first if not already in context — it is the source of truth for the rules above.
When unsure whether a comment carries real non-local context, keep it and flag it as "keep (verify)" rather than deleting.