huashu-flash skill
A skill that speeds up a website: measure, regression guards, a loop of small edits and paired comparison.
- 112 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The skill edits site code, makes commits and runs browser measurements.
- Releasing to the site happens only after user approval.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/huashu-flashDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill edits site code, makes commits and runs browser measurements.
- Releasing to the site happens only after user approval.
This is third-party code. Review the repository files before installing.
What it does
huashu-flash, "Flash", turns the method from the Anthropic engineering post "How we made claude.ai 3x faster in two weeks" into eight steps an agent runs on your site. The steps are: write a kickoff brief on key pages and what counts as usable; define "open to usable", excluding skeletons and placeholders; take lab baselines with cold cache, a slow network and p75 over ten runs, adding real-user data; set regression guards before touching code; find a countable metric and prove it tracks user feel; run a loop of one change, one commit, revert if not faster; leave courage, taste and direction to the human; compare versions with paired runs and get approval to ship. The `scripts` folder holds `bench.py` for measurement (it needs Playwright) and `ratchet.py` for locking in gains.
Who it is for. Site owners and frontend developers who ask an agent to speed up pages and want regression guards and a fair comparison.
Good fit when
- A site opens slowly and you need to see where first-screen time goes.
- Before optimizing you need a set of checks so speedups do not break design or SEO.
- You need to compare versions before and after in pairs, alternating runs.
- Pages are heavy with big images and scripts and you want to remove them one commit at a time.
Not a fit when
- You need server or database optimization.
- Other heavy tasks run on the machine, so measurements taken at different hours will not be comparable.
- The site cannot be changed without sign-off and no human is available to approve.
Example request
Speed up the pricing page: write a kickoff brief, take cold-cache baselines and make edits one at a time, reverting any that give no gain.Limitations
The skill texts are written in Chinese, so the agent must translate the guidance. `bench.py` needs Playwright, installed with `pip install playwright`. Measuring production from a home machine through a proxy distorts time to first byte, so take a reference site's number first. Free hosting quotas for image conversion can run out, so prepare images at build time. The skill does not promise a tenfold speedup on every site: the result depends on what is slow.
How to disable. Delete the `huashu-flash` folder from the agent skills directory. Commits made to the site stay in git history.
FAQ
How do I install huashu-flash?
Run `npx skills add alchaincyf/huashu-flash`. You can also give the repository link to Claude Code or Codex and ask it to install the skill.
What does open-to-usable mean?
The moment a user can work with the page. Skeletons and placeholders do not count, so you pick the metric in advance with the agent.
What is the ratchet for?
It records the gained value in the baseline and allows only improvements: a regression fails the check.
How are versions compared?
With paired runs: the two variants run alternately, because measurements taken at different times are unreliable when the machine is loaded.
Who decides on shipping?
The user. The skill leaves courage, taste and direction to the human, and the site is released only after your approval.
Discussion
No questions or reviews yet.
Related
Skills56 JavaScript snippets for web performance: the agent runs them through skills, MCP or CLI and reasons over the result
A skill methodology for reproducing websites, from a static page to heavy WebGL, built on real source code rather than agent-hallucinated code
The /taste skill turns a URL into design tokens and trade-off principles, with ready rules for Cursor and Claude Code
A set of skills for web quality audits: performance, Core Web Vitals, accessibility, SEO and best practices from Lighthouse
A free Claude skill that writes a build spec, picks cinematic media, builds a landing page and checks it with screenshots
A skill rebuilds a UI from a site, screenshot or screen recording and scores the copy on layout, style, pixels and motion.
README in short
The README reports measurements on four real sites: large gains on two and modest ones on two already optimized. The pitfalls already hit are written into the skill so the agent does not repeat them.
README badge
Original SKILL.md
--- name: huashu-flash description: 闪电.skill：照 Anthropic 两周把 claude.ai 提速3倍的方法给网站或 App 提速。定义「打开到能用」、测基线、找能数的指标并证明它跟体感挂钩，逐步优化，用棘轮锁住只许变好。说网站太慢、提速、性能优化、首屏、秒开时用。 metadata: author: "花叔" homepage: "https://github.com/alchaincyf" repository: "https://github.com/alchaincyf/huashu-flash" version: "1.0.0" --- # 闪电.skill · huashu-flash > 《疯狂动物城》车管所里那只树懒，名字叫「闪电」。你的网站可能也是。 ## 你是谁 你是那个性能频道里的 Claude，一个数字狂魔：任何能数出来的东西，你都能把它往下压。但你知道能数不等于该数，所以每个数字先证明自己跟用户的体感挂钩，再去爬它。 你交付的不是「感觉快了」，是一张前后对比表：同一口径、同一时段、交替配对测出来的 p75，外加一套只许变好的棘轮，保证这次赢下来的东西以后不会悄悄流失。 你也清楚自己的毛病：默认太保守。发现问题想先记成待办，估时往宽了报，一达标就想停。这里的规矩是，护栏够结实就大胆改，目标不是终点。 ## 这套方法从哪来 Anthropic 2026年9月的工程博客《How we made claude.ai 3x faster in two weeks》：两周、一个 Slack 频道、Claude 在每个讨论串里干活，合入3000多个改动，零线上事故零回滚，13项核心指标 p75 几何平均快3.1倍。原文：https://claude.dev/blog/how-we-made-claude-ai-faster/ 它的核心就一句：**只要 Claude 能把一件事量出来，它就能把这件事变快。** 测量以前是第零步（加埋点、等数据、再理解问题），现在是爬山的第一步。所以杠杆最高的事是找到更多能数的东西，并且确认你数的就是用户感受到的。 ## 八个步骤（按顺序，每步有产物） 所有产物放在项目里的 `perf/` 目录。 ### 1、写开工简报 先问用户，只问他才知道的： - 核心页面或核心操作是哪几个（拿不准就按流量排，能查统计就查） - 对每个操作，什么叫「能用」（能打字？能点主按钮？主要内容可见？） - 哪些东西不能动：视觉、功能、SEO 字段、第三方统计 - 目标：用户给了就用用户的；没给就定「打开到能用的 p75 降一半」 - 上线谁批准、怎么回滚 照 `references/brief-template.md` 写成 `perf/BRIEF.md`。 ### 2、定义「打开到能用」