Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Updates README.md and README.zh-CN.md to reflect the project's current state. Use this skill whenever the user asks to "update the README", "sync the docs", "update documentation", "reflect latest changes in README", or wants both the English and Chinese READMEs to match the current project.
.claude/skills/fradser-update-readme/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 644% | 0% |
| case-04 | ✓→✗ | ▼ Worse | 75% | 0% |
| case-05 | ✓→✗ | ▼ Worse | -46% | 0% |
| case-07 | ✓→✗ | ▼ Worse | -22% | 0% |
| case-17 | ✓→✗ | ▼ Worse | -79% | 0% |
Keep README.md (English, primary) and README.zh-CN.md (Simplified Chinese, secondary) in sync with the project's actual current state. Both files must be accurate, complete, and consistent with each other.
Every README must open with this exact structure — adapt the badges, project name, and description to the project at hand:
markdown# <Project Name>  [](<link>) [](<link>) **English** | [简体中文](README.zh-CN.md)
For README.zh-CN.md, reverse the language toggle:
markdown# <Project Name>  [](<link>) [](<link>) [English](README.md) | **简体中文**
Choose badges that reflect what is genuinely true about the project. Common types: CI status, license, language/runtime version, package registry version, code coverage. Use shields.io static badges when there is no live endpoint. Prefer reference-style Markdown links for badge rows with more than two badges — it keeps the source readable. Broken or always-failing badges are worse than no badges; only include ones that are maintained.
Before writing anything, read the actual project state:
Write from ground truth, not from memory or assumption.
Use references/template.md as the structural starting point. Adapt section names to the project's domain — the template shows the skeleton, not the words.
Section order (omit what doesn't apply, don't add sections just to fill space):
For READMEs over ~300 lines, add a Table of Contents after the one-liner.
The README is a directory, not a tutorial. Keep each component description to one or two sentences. Installation commands must be copy-pasteable — exact commands, no ambiguity. Always use fenced code blocks with a language tag ( bash , json ).
Write with a human voice. README prose is some of the most AI-trope-prone writing that exists — it tends to accumulate "robust", bold-first bullet points, and "serves as" constructions without anyone noticing. Specific things to watch for:
A useful test: read each sentence aloud. If it sounds like promotional copy, rewrite it as a plain statement of fact.
Translate the English README faithfully. Rules:
[English](README.md) | **Simplified Chinese**.Before writing either file, verify:
Then scan the draft for writing issues:
Write README.md first, then README.zh-CN.md. Use the Edit or Write tool — do not output the content as a code block in the conversation.
After writing, briefly confirm what changed (e.g., "Added update-readme skill, removed stale apple-events prerequisite note").
references/template.md — README structure template (load when drafting)| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 2,362 | 8,334 | +253% | 1 | 1 | 0% | 328 | 1,637 | +399% | 0 | 0 | — |
case-02 | fail→fail | 4,955 | 3,620 | -27% | 1 | 1 | 0% | 212 | 1,604 | +657% | 0 | 0 | — |
case-03 | fail→fail | 4,241 | 3,763 | -11% | 1 | 1 | 0% | 184 | 1,620 | +780% | 0 | 0 | — |
case-04 | pass→fail | 4,958 | 4,800 | -3% | 1 | 1 | 0% | 930 | 1,632 | +75% | 0 | 0 | — |
case-05 | pass→fail | 13,920 | 4,186 | -70% | 1 | 1 | 0% | 3,010 | 1,632 | -46% | 0 | 0 | — |
case-06 | fail→fail | 4,088 | 7,318 | +79% | 1 | 1 | 0% | 769 | 1,848 | +140% | 0 | 0 | — |
case-07 | pass→fail | 12,189 | 5,123 | -58% | 1 | 1 | 0% | 2,222 | 1,739 | -22% | 0 | 0 | — |
case-08 | fail→fail | 7,318 | 3,919 | -46% | 1 | 1 | 0% | 1,327 | 1,659 | +25% | 0 | 0 | — |
case-09 | fail→fail | 10,840 | 4,599 | -58% | 1 | 1 | 0% | 1,651 | 1,712 | +4% | 0 | 0 | — |
case-10 | fail→fail | 5,373 | 5,422 | +1% | 1 | 1 | 0% | 811 | 1,828 | +125% | 0 | 0 | — |
case-11 | fail→fail | 5,408 | 4,175 | -23% | 1 | 1 | 0% | 934 | 1,655 | +77% | 0 | 0 | — |
case-12 | fail→pass | 6,259 | 27,874 | +345% | 1 | 1 | 0% | 948 | 7,050 | +644% | 0 | 0 | — |
case-13 | fail→fail | 7,987 | 3,994 | -50% | 1 | 1 | 0% | 1,414 | 1,687 | +19% | 0 | 0 | — |
case-14 | fail→fail | 5,782 | 4,980 | -14% | 1 | 1 | 0% | 1,074 | 1,744 | +62% | 0 | 0 | — |
case-15 | fail→fail | 13,252 | 3,335 | -75% | 1 | 1 | 0% | 2,693 | 1,581 | -41% | 0 | 0 | — |
case-16 | fail→fail | 6,031 | 4,908 | -19% | 1 | 1 | 0% | 960 | 1,660 | +73% | 0 | 0 | — |
case-17 | pass→fail | 43,453 | 4,959 | -89% | 1 | 1 | 0% | 8,211 | 1,715 | -79% | 0 | 0 | — |
case-18 | fail→fail | 9,029 | 4,611 | -49% | 1 | 1 | 0% | 1,380 | 1,694 | +23% | 0 | 0 | — |
case-19 | fail→fail | 7,290 | 6,173 | -15% | 1 | 1 | 0% | 1,243 | 1,756 | +41% | 0 | 0 | — |
case-20 | fail→fail | 26,960 | 5,982 | -78% | 1 | 1 | 0% | 3,921 | 1,701 | -57% | 0 | 0 | — |
case-21 | fail→fail | 4,443 | 3,714 | -16% | 1 | 1 | 0% | 833 | 1,605 | +93% | 0 | 0 | — |
case-22 | fail→fail | 17,156 | 5,750 | -66% | 1 | 1 | 0% | 2,622 | 1,870 | -29% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 1 counted toward the lift figure. The other 21 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of -14 percentage points is the difference between those two pass rates over the 1 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.