AGENT SKILL

refresh-readme

One repeatable pass over README.md — the front door — that keeps ONLY the most important points current and honestly framed. Runs the freshness auditor (tools/readme_freshness_audit.py), turns each FAIL into a required edit and each WARN into a judgment call, applies the three front-page laws (SOTA-vs-us-never-naive, 6th-grade/Feynman-but-accurate, wide-audience), re-stamps the readme-verified marker, and commits ONLY README.md (+ the tool) by explicit path. Use after a release / version bump, after a headline number changes, when a link dies, or on a /loop cadence.

License 许可
Apache-2.0
最近更新
2026-11-15
Stars
★ 48
Forks
16

01 前言 Front Matter

name
refresh-readme
metadata
{ "canonical": "../../../.claude/skills/refresh-readme/SKILL.md", "generated-by": "fak project-assets sync" }
description
One repeatable pass over README.md — the front door — that keeps ONLY the most important points current and honestly framed. Runs the freshness auditor (tools/readme_freshness_audit.py), turns each FAIL into a required edit and each WARN into a judgment call, applies the three front-page laws (SOTA-vs-us-never-naive, 6th-grade/Feynman-but-accurate, wide-audience), re-stamps the readme-verified marker, and commits ONLY README.md (+ the tool) by explicit path. Use after a release / version bump, after a headline number changes, when a link dies, or on a /loop cadence.

03 SKILL.md 正文

refresh-readme — keep the front door correct, honest, and small

What this does. README.md is the one surface read cold by everyone — adopter, reviewer, skeptic — and the one most likely to rot: a link dies, a version pin lags VERSION, a headline number drifts from BENCHMARK-AUTHORITY.md, a "we beat naive" claim creeps back into the lead. Every other claim surface here has a checking layer (a memory-recall audit for memories, a closure audit for issue closes, BENCHMARK-AUTHORITY for numbers). This is the README's. It makes "keep only the most important points on the front page, framed honestly" a repeatable pass, not a one-time edit that decays the moment the person who did it moves on.

The shape: run the auditor → fix every FAIL → weigh every WARN → apply the three laws → re-stamp → commit ONLY the README lane.


The three front-page laws (the durable policy this pass enforces)

These are not style preferences — they are standing rules for the front page, and the auditor checks each one. Internalize them; the README is graded against them every pass.

  1. SOTA-vs-us, never naive. Every headline number on the front page compares fak against the best already-shipped alternative, not a strawman. "~4× vs a tuned warm-cache stack" leads; "~60× vs naive" does not — beating naive is easy, and leading with it invites the "you're fighting a strawman" dismissal. A naive number may appear as an honest aside that says it's not the headline; it may never be the bolded lead. → auditor naive_baseline FAIL.

  2. 6th-grade / Feynman voice, but still accurate. The front page is the audience-widening surface; the deep-dive links are where the jargon lives. On the first screen, lead with the plain-English idea, then name the term in parens — "a scratchpad of the work-so-far (the KV cache)", not "the KV cache". Explain by concrete example, the Feynman move, before reaching for the abstraction. Accuracy is not negotiable — simplify the words, never the claim. Every acronym gets a parenthetical on first use. → auditor jargon_density ADVISORY.

  3. Wide-audience appeal. The first screen gives each reader a foothold: the skeptic (what's real / what's not), the security lead (the lock, not the screener), the perf engineer (the reuse win + its fences), the casual reader (the 2-minute no-key demo). If a section serves only one audience, ask whether it belongs on the front page or behind a "Go deeper" link.

And the size law that wraps all three: the front page holds only the most important points. Before adding anything, ask — would this earn its place if the page could hold only ten things? If not, it belongs in a linked topic doc, not on the front page. Detail flows OUT to docs/ and the "Go deeper" table; the front page stays small.

The size law now has teeth (it used to be aspirational). Five of the six substance checks reward adding an affordance — put fak guard up top, add a speed number, add a hero result, add a persona router. Optimized to those alone, the page only ever grows: the git log shows it halved on 2026-07-01 and immediately regrew, one well-meaning "surface X on the front page" commit at a time. front_page_focus is the counterweight — a line budget, a section-count budget, and a single-lead rule (the one-binary/syscall pitch stated once in the preamble, not restated three times before the reader reaches a section). It feeds the composite score and the readme_debt gate, so a bloated page can no longer score an A. The two forces now both live in the tool; concision is checked, not just hoped for.

The three anti-regrowth rules this pass enforces:

  1. Retire before you add. The page is at its section budget by design. To add a section, fold or cut one first — do not bump FRONT_PAGE_SECTION_BUDGET to make room. Bumping a budget to silence the warning is the failure mode; bump it only when the page has genuinely, durably earned the slot, and say why in the commit.
  2. One lead, not three. The pitch is derived from docs/adoption/pitch-ladder.md rung 1 and appears once above the first ## . A second "and also, fak is a binary you put in front of your agent" paragraph is the confusion single_lead catches — collapse it into the one lead.
  3. Detail flows OUT, never back IN. The overflow sink is docs/README-legacy.md. Narrower-audience or deep-dive material moves there and earns a link; it does not migrate back onto the front page. If a section keeps wanting to come home, that is a signal to write a topic doc, not to re-inline it.

Step 1 — Run the auditor (it builds your work-list)

From the repo root:

python tools/readme_freshness_audit.py --json > readme-before.json
python tools/readme_freshness_audit.py            # human rendering of the same evidence

readme-before.json is the mandatory first artifact. Quote every binding FAIL check name and its evidence before editing. Retire all binding FAIL rows before advisory prose polish; an advisory unglossed term such as KV cache must not displace a red guard_prominence check. The current README product defect belongs in its dedicated README issue, not in this skill-only adjudication.

It checks, and exits non-zero on any FAIL:

check fires on severity
links a local Markdown link whose target is missing on disk FAIL
version_pins a vX.Y.Z string behind the VERSION file FAIL
naive_baseline a bolded headline that LEADS with a "naive" baseline (law 1) FAIL
headline_authority a bolded multiplier not mirrored in BENCHMARK-AUTHORITY WARN
freshness_stamp the readme-verified marker absent or older than 14d WARN
jargon_density first-screen expert terms with no plain gloss nearby (law 2) advisory
front_page_focus the page busts the line/section budget or restates the lead ≥3× (the size law) debt

FAIL = a required edit. WARN = a judgment call. ADVISORY = a nudge. Voice (jargon) is never a hard gate — plain-language is writing judgment, not a mechanical rule.

Step 2 — Fix every FAIL

Work in the exact binding order quoted from readme-before.json; do not start WARN/ADVISORY polish or the separately issue-tracked README product fix while any FAIL remains.

  • dead link — the target moved or was deleted. Repoint it to the current path, or drop the link if the doc is gone. (Don't invent a path; verify it exists.)
  • stale version pin — bump it to match VERSION. A deliberate forward range (v0.31.x) on the current minor is fine and passes.
  • naive-lead headline (law 1) — invert it. Put the SOTA comparison in the bold lead; demote the naive number to a plain-prose aside that names itself as not-the-headline, or cut it.

Step 3 — Weigh every WARN, apply laws 2 & 3

Enter this step only after all binding FAIL rows are green.

  • headline_authority WARN — a front-page number isn't traceable to BENCHMARK-AUTHORITY. Either it's stale (fix it to the authority figure) or it's a number that shouldn't be on the front page at all (an untraced claim). Reconcile against the authority doc; never invent a number to match.
  • jargon ADVISORY — for each flagged first-screen term, add a one-clause plain gloss the first time it appears (law 2). Don't touch the deep-dive links.
  • read the first screen as each audience (law 3) — does the skeptic, the security lead, the perf engineer, and the casual reader each get a foothold in the first screen? If a point serves none of them, it's a candidate to move behind a link (the size law).

Step 4 — Re-stamp the freshness marker

After the page is correct, update the stamp near the top of README.md to today's date and the current VERSION:

<!-- readme-verified: YYYY-MM-DD vs VERSION X.Y.Z + BENCHMARK-AUTHORITY · process: tools/readme_freshness_audit.py + /refresh-readme -->

This is the freshness contract: it's how the next reader (and the next audit) knows the page was checked against reality, and when. Re-run the auditor — it should now be green (exit 0).

Step 5 — Commit ONLY the README lane, by explicit path

On a shared tree, HEAD can move under you and peers may have dirty paths. The commit discipline:

  • Stage by explicit path, never git add -A — commit your README, not a peer's half-written code:
    fak sync reconcile --apply
    fak commit --path README.md -F <msgfile> [--push]
    fak sync push
    
  • Doc-only diff → docs(readme): … subject, NOT fix(/feat(. A code-effect prefix on a docs-only diff overclaims — keep the prefix honest to what changed.
  • On Windows, pass the message via a file (-F), not a here-string — native exe arg passing mangles multi-line quotes.
  • If a peer's MERGE_HEAD is set (cannot do a partial commit during a merge): wait for it to clear — don't abort or work around it. Markdown self-heals; re-try the pathspec commit once MERGE_HEAD is gone.
  • Stay on the trunk (main) — never branch or worktree to dodge a dirty/diverged tree. Push promptly via fak sync push (or --push).

If a release just happened, this pass typically only needs Step 4 (re-stamp to the new VERSION) — the auditor catches the bump immediately.


This is already a durable loop — you don't have to remember to run it

The front page does not rely on someone thinking to check it. The checking half runs on a fixed cadence, and this skill is the acting half it hands work to:

  1. The auditor is a registered gardening loop. readme-freshness is an enabled entry in tools/control_pane.loops.json, wired to python tools/readme_freshness_audit.py --json. It sits alongside the other front-door checking layers (memory-recall-audit, docs-scorecard, …).
  2. It runs daily, unattended. .github/workflows/garden.yml runs fak garden on a daily cron (06:23 UTC), and fak garden folds the fleet loop-audit — which runs every enabled gardening loop once, this one included. So the README is re-checked against the filesystem, VERSION, and BENCHMARK-AUTHORITY every day, whether or not anyone asked.
  3. The loop itself is watched. The gardening-loops-audit meta-loop flags readme-freshness as BROKEN if its status command ever stops running — the loop can't silently rot into a no-op.
  4. This skill is the acting half. When the daily audit returns ACTION (ok:false — a FAIL), that verdict names /refresh-readme as the runbook. You (or a supervisor turn) run the pass below; the loop closes.

So "a durable README loop that routinely gardens the front page" already exists end to end: daily audit → ACTION verdict → this skill → commit the README lane. The section below is what you do when the loop hands you an ACTION (or when one of the triggers fires early).

When to run this

  • After a /release or any VERSION bump (the stamp + any pin go stale at once).
  • When a headline number changes in BENCHMARK-AUTHORITY.
  • When a doc the README links to moves or is renamed.
  • When the daily readme-freshness loop (above) returns ACTION, or on a manual /loop cadence to garden the front page ahead of the daily tick.

The auditor is read-only; this skill's only writes are README.md and re-running the tool. It never edits a deep-dive doc — the front page is the only surface in scope.

04 同仓库更多 Skills