
Human Review
Skill用 npx human-review 在浏览器里打开 agent 写的 HTML、Markdown 文件或本地开发页面,用户可直接改字、选中文字留评论,点发送后 agent 一次收到全部修改与评论并应用到源文件。适合评审方案、报告、草稿与落地页。
name: human-review
description: Open an HTML file, Markdown file, or localhost page in the browser so the user can edit text directly and leave comments on specific parts, then send all edits and comments back to you. Use after writing or updating something the user will read — specs, plans, reports, newsletter drafts, landing pages, slide decks, and locally running web pages.
human-review
The user reviews your HTML, Markdown, or localhost page in a real browser: they fix small things by typing, select anything to comment on it, and send you the whole batch at once.
Markdown files open rendered. Their quotes and edits reference the rendered text, and the file itself is never touched — apply every change to the Markdown source, keeping its formatting syntax.
The loop
-
Write or update the HTML or Markdown file, or start the local page being reviewed.
-
Open it for the user:
npx -y human-review path/to/file.htmlFor a page served by a local development server, open the real route instead of recreating it as a separate HTML file:
npx -y human-review http://localhost:3000/wiki -
Wait for feedback. This command blocks until the user hits Send in the browser, then prints their batch and exits:
npx -y human-review poll path/to/file.htmlThe command exits only when the user clicks Send or closes the review. There is nothing to re-run, no interval to poll on, and no
--timeoutto add. It survives the local server restarting, and feedback is saved even if the poll dies, so nothing is ever lost. How you wait depends on your harness:-
Claude Code: run it with
run_in_background: trueand end your turn. Claude Code wakes you with the output the moment the command exits. -
Codex, Cursor, and everything else: run it in the foreground, inside your active turn, and stay on it until it prints
feedbackorclosed. Do not detach it or start it as a background session: nothing wakes you when a detached command finishes. While the review is active:- If the user sends a message, answer it as commentary and immediately
resume the foreground poll in the same turn. Do not send a final
response until the poll returns
feedbackorclosed— a final response ends the turn and kills the wait. - If your shell tool caps command duration, pass
--timeouta little under the cap and run bounded polls back to back in the same active turn until one returnsfeedbackorclosed.
Know the limit: this is reliable only while your turn stays active. A turn that has already ended is not woken when the user hits Send; the user has to message you, and you then run
statusandpollto pick the batch up. There is no integration that resumes an ended task when the poll exits. - If the user sends a message, answer it as commentary and immediately
resume the foreground poll in the same turn. Do not send a final
response until the poll returns
If it prints
{"status":"closed"}, the review is over: the user ended it, closed the tab, or never had one open (reasonsays which). Stop and do not start another poll.unsentcounts feedback they left behind; if it is not zero, tell the user in one line that it is kept and they can restore or discard it next time.{"status":"superseded"}means a newer poll of yours owns the wait — stop this one silently.{"status":"timeout"}only appears after 12 hours; runstatusand start the wait again if the review is still open. -
-
Apply what comes back, then start the next background poll.
--ackclears the batch you just handled:npx -y human-review poll path/to/file.html --ack
Repeat 3–4 until the user says they are done.
Not sure whether feedback is already waiting — say, at the start of a new turn with no poll running? This answers instantly without blocking:
npx -y human-review status path/to/file.html
It prints {"status": "feedback-waiting"} when a batch is ready for a poll,
plus counts of unsent comments and edits still in the browser.
What you get
One batch covers every page the user visited, grouped by file or localhost URL.
{
"status": "feedback",
"pages": [
{
"file": "/abs/path/to/page.html",
"edits_saved": true,
"comments": [
{ "id": "c_1", "kind": "selection", "quote": "the exact text they selected",
"anchor": { "prefix": "...", "quote": "...", "suffix": "..." },
"feedback": "what they want changed" }
],
"edits": [
{ "label": "Problem body", "kind": "edited",
"before": "the original wording",
"after": "their exact new wording",
"after_html": "their exact new wording with <strong>formatting</strong>" }
]
}
],
"overall_note": "feedback not tied to any one page"
}
Rules
editsare changes the user already made.afteris their exact wording — carry it across verbatim and never revert it. If the HTML was generated from something else (MDX, Markdown, a template), applyafterto the source too, or their fix disappears on the next build.edits_saved: truemeans those edits are already in the file on disk. Plain HTML files autosave as the user types, so your copy of the file is stale. Re-read the file before touching it and make targeted changes only; never regenerate it from what you wrote earlier, or their work disappears.edits_saved: false(Markdown, localhost pages, self-rendering HTML) means the edits exist only in this batch — apply them to the source yourself.- An edit with
kind: "deleted"means the user removed that whole block: delete it from the source too, without asking why. - An edit marked
truncated: truehad its text cut at 200k characters; read the block from the page itself rather than fromafter_html. - When
before_html/after_htmlare present, the user changed formatting, not just words — bold, italic, underline, links. Use the HTML version to carry the formatting into the source, translated to its syntax (e.g.<strong>→**in Markdown/MDX). - A page with
kind: "url"was edited directly in the review UI. Itsfileandurlfields name the localhost route, not a writable file. Find the matching project source (such as MDX, TSX, or a template), apply every edit and deletion there, then acknowledge so the route reloads. Never write the rendered HTTP response back into the app. - When an edit's
after_htmlcontains<img src="assets/...">, the user pasted an image: the file already exists in anassets/folder next to the reviewed file. Keep that relative path — in Markdown, reference it as. Never regenerate or inline the image. - On a localhost page, a pasted image arrives under
staged_assets. Copy its localpathinto the app's appropriate asset folder, replace the temporary preview URL inafter_html, and preserve the image at the user's insertion point. Never leave the temporary preview URL in source. - An edit with
kind: "moved"means the user relocated that whole block. Reposition it in the source without rewriting its content: it now sits right after the block whose text starts withmoved_after, and right before the block whose text starts withmoved_before(both are clipped to 90 characters and may end in…). An emptymoved_aftermeans it is now the first block in its container. - Find each comment by its
quote. It is the rendered text the user selected, so in Markdown or templated HTML it may span formatting syntax or tags;anchor.prefixandanchor.suffixgive the surrounding text to disambiguate. kind: "element"points at a whole block, soquoteis its label, not body text.- Copy any
staged_assetsfiles before you ack:--ackdeletes them. - A batch with only an
overall_notehas an emptypagesarray. - Fix every page in
pages, not just the first. - Do not write a reply. There is no chat. The user sees your work when the page reloads, which happens on its own the moment you save the file.
Better edit labels (optional)
Name the sections you author and the user's edit list uses your names instead of guessing from the DOM:
<p data-block="Problem body">…</p>
<div data-container="Metrics callout">…</div>
data-block names a region for the edit list. data-container also makes the block
clickable as a comment target.
相關資源

Awesome Claude Skills (Composio)
SkillClaude 技能精选索引,仓库内自带约 30 个实用技能(简历定制、内容调研写作、文件整理、会议分析等),另有 800 多个经 Composio 连接第三方应用的自动化技能,需接入 Rube MCP 并授权账号。
ComposioHQ/awesome-claude-skills79.8k
Doc Co-Authoring
Skill引导用户分三个阶段协作撰写提案、技术规格与决策文档:收集上下文、逐节迭代完善、用不带上下文的新会话做读者测试,提前发现读者读不懂的地方。Anthropic 官方技能。
anthropics/skills197k
Writing Guidelines
SkillVercel 官方技能:联网拉取 Vercel 写作手册的 80 多条规则审查文档与文案,覆盖语气、结构、内容类型、代码示例与排版,以 file:line 格式列出问题。规则面向英文文档。
vercel-labs/agent-skills34.5k
Yunshu SkillsHub
Skill云舒整理的中文工作流技能合集,近 40 个技能覆盖需求池、版本规划、PRD 与测试用例、页面设计探索、案例调研、写作与周报,帮助产品经理把想法推进到交付。
yunshu0909/yunshu_skillshub790
X Article Publisher
Skill把写好的 Markdown 长文一键排进 X(Twitter)Articles 编辑器:自动处理标题、加粗、链接、封面和正文配图位置,表格与 Mermaid 图转成图片,只保存草稿、不会自动发布。需要 X Premium Plus 账号、Playwright MCP 和 Python(macOS / Windows)。
wshuyi/x-article-publisher-skill875
Mintlify
PluginMintlify 官方插件,已收录于 Anthropic 官方插件市场:提供 Mintlify 文档站的组件、docs.json 配置、导航与 API 文档写法参考,并接入 Mintlify 文档搜索与站点管理服务,帮技术写作者和客服团队搭建、更新产品文档与帮助中心。需要已有 Mintlify 文档项目,管理功能首次使用要在浏览器授权 Mintlify 账号。
mintlify/mintlify-claude-plugin7