本文へ移動
cccskills
無料GitHub で公開

docs-reader-review

Use when a docs page or section has been written or rewritten and is about to be handed over, when the operator says "reader review", "does this read like a human wrote it", "too much jargon", "plain language", or when a docs brief asks for a review before a pull request.

インストール方法を見る

含まれるファイル(8)

  • SKILL.md9.5 KB
  • references/ai-writing-signs.md9.8 KB
  • references/banned-terms.md5.5 KB
  • references/explain-not-state.md5.2 KB
  • references/reader-persona.md1.4 KB
  • scripts/check-ai-signs.sh4.8 KB
  • scripts/check-plain.sh2.3 KB
  • scripts/check-staccato.py4.2 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Docs Reader Review

Run a page through the eyes of the reader it is for, before anyone else sees it. The biggest complaint about the Prisma docs is that pages read as if a model wrote them: vocabulary from the source code, mechanism instead of action, several ideas per sentence. A fact review does not catch this, because every such sentence is true. This skill catches it.

Use it after docs-writer (or any other writing pass) and after the fact review. Nothing ships with an unresolved mark.

Pre-conditions

  1. The page exists on disk and every claim on it has already been checked against the source. This skill reviews comprehension, not truth. If facts are unverified, do that first.
  2. You know who the reader is. Default: a developer who has used the previous major version for two years and has never seen this version's source code, release notes, or internal vocabulary. If the page targets someone else, write that reader down before starting.

Workflow

1. Run the banned-term check

Run .claude/skills/docs-reader-review/scripts/check-plain.sh (path from the repository root) on every changed page. It fails on the words in references/banned-terms.md, which lists source-code vocabulary and the plain words to use instead, and it ignores code blocks and inline code. (Machine-writing vocabulary such as "robust" or "leverage" is a different list, checked in step 1c.) Replace every hit before going further. Do not argue that a term is fine in context. The one exception: if the sentence itself defines the term in plain words, add {/* plain-language:defined */} to that line; the checker skips it and the reader review will confirm the definition landed. Terms the list marks "(unexplained)" are not checked by the script; the reviewer catches them.

1b. Run the staccato check

Run .claude/skills/docs-reader-review/scripts/check-staccato.py on every changed page. It flags paragraphs with three or more consecutive sentences under nine words, the usual signature of a page whose long sentences were split without keeping the connectives. Read each flagged paragraph aloud. Rejoin sentences that are cause and effect, contrast, or condition and result with "because", "so", "but", "while", or a colon, and leave apart the ones that are separate ideas. The script only finds the worst runs; a paragraph of eleven-word sentences that all land the same way is still staccato, so read the whole page for rhythm, not only the hits. Reference entries that are fragments by design (a Payload: line, a one-line table note) are not prose and do not count.

1c. Run the AI-signs check

Run .claude/skills/docs-reader-review/scripts/check-ai-signs.sh on every changed page. It flags the signs of machine-written prose that a regex can catch: the over-used vocabulary (crucial, robust, seamless, leverage, showcase, "Additionally,"), "serves as" in place of "is", "not just X but Y", "it's important to note", "In summary", chat text pasted into the page ("Here's an overview of", "I hope this helps", unfilled [placeholders]), em dashes, curly quotes, --- breaks between sections, and Title Case headings. references/ai-writing-signs.md explains each one and gives the plain replacement; it is adapted from Wikipedia's Signs of AI writing. Fix every hit. The signs a regex cannot catch (participle tails like "..., ensuring your app scales", the rule of three, bold-label lists everywhere, the same thing renamed in every paragraph, features described by their importance instead of their behavior) are in the same file; read it once before the reader round, because the reviewer reports them as "sounds like marketing" or "could not restate" without naming the pattern.

1d. Read for explanation, not statement

Read references/explain-not-state.md before fixing anything. It names the habit that a plain-language pass and a word budget both produce: true, short sentences that state a fact and never say what it means for the reader. The staccato check flags the mechanical signs (counting lead-ins like "Four things change it:", fragment openers like "One name is special."), but most of the work is the desk test in that file: read each paragraph aloud as if to a colleague, and add the sentence you would say out loud. A page may grow when it gains explanation. Word budgets are for repetition only.

2. Dispatch the reader

Hand the page to a fresh reviewer that has no memory of writing it, using references/reader-persona.md verbatim as its instructions. The reviewer reads the page once, top to bottom, with nothing else open, and reports:

  • every sentence it could not restate in its own words, with what stopped it;
  • every word or phrase it had to guess;
  • every place it asked "so what do I type?" and the page did not say;
  • every place the page explains how the tool works inside when the reader only needed what to do;
  • whether it could complete the page's purpose after one reading, and what it would still not know.

The reviewer must not look anything up. Its confusion is the data.

3. Fix every mark

For each reported sentence:

The reviewer saidDo this
a word it had to guessreplace with the plain words from references/banned-terms.md, or define it in that sentence
too many ideas in one sentenceone idea per sentence, keeping the connective (because, so, but) that ties related facts together; a row of clipped one-clause sentences is its own failure, not a fix. A table cell with three ideas becomes a note under the table
"so what do I type?"add the command or the code, or the link to the page that has it
mechanism instead of actiondelete the mechanism, keep what the reader does and what they see
a reference it could not resolve ("the plan", "the ref", "spec")show the thing, or name where it comes from
a term used before the page defines itmove the definition to the first use
"reads like marketing", "sounds generated"find the pattern in references/ai-writing-signs.md (participle tail, puffed significance, negative parallelism, rule of three) and replace it with the specific fact

Do not add content the page's purpose does not need. Do not argue with the reviewer. If a mark seems wrong, the sentence still confused a reader; rewrite it anyway.

Facts stay fixed. A reader round changes wording, order, and examples. It never changes what the page claims. When a reviewer's confusion can only be resolved by a fact you do not have (which command to run, what a value is, what happens on failure), do not invent one to make the sentence smooth: look it up in the source, or mark it Q1, Q2 for the operator and leave the sentence honest. Rewrites that read well and say something false are the worst outcome this skill can produce.

4. Repeat

Dispatch a fresh reviewer again on the fixed page. Repeat until a pass reports no sentence it could not restate and no word it had to guess. Two rounds is normal. One round is suspicious.

5. Re-check the facts

Reader rounds rewrite sentences, and rewriting drifts meaning. Before handing over, check every claim in the final text against the source again: each command, flag, file path, return value, and "what happens if". Anything the rounds introduced that the source does not support is reverted to what the source says, even if it reads worse.

6. Hand over

Only now does the page go to the operator or into a pull request. Say in the handover how many review rounds ran, what the last round still flagged, and any Q marks left for the operator. Remove all Q marks from the page itself before handover; carry them in the handover note.

Rationalizations that do not hold

ExcuseReality
"The term is the correct name for the thing."Correct for whom? If the reader has not met it, it is noise. Define it or drop it.
"The fact review passed, so the page is fine."The fact review checks truth. Every jargon sentence is true.
"Reference pages are for experts; they can take the vocabulary."The reader of a reference page is looking something up mid-task. They know their own code, not the tool's internals.
"Explaining the mechanism helps the reader understand."Explain it only if the reader must know it to act. Otherwise it is the sentence they skip and the one that loses them.
"There is no time for a second review round."The second round is where the remaining marks are. A page shipped after one round is the page that gets the complaint.
"I already know what a reader would say."You wrote it. You cannot read it cold. Dispatch the reviewer.
"It is only a table cell."Table cells are where three ideas get packed into one line. They are the first thing to check.
"The reviewer could not follow it, so I explained what must be happening."You just invented a fact. If the source does not say it, the page does not say it. Mark it for the operator.
"After six rounds it reads perfectly."Reads perfectly and says what? Run the fact check on the final text; smooth prose is where drift hides.

Output

The reviewed page, with no marks left, plus the count of review rounds. Do not carry the reviewer's report into the page.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use when the operator wants a hero or meta image for a Prisma blog post; asks to create or generate a blog hero, cover, social card, Open Graph, or YouTube image; mentions cover art, a blog thumbnail, cover.svg/hero.svg/meta.png; references content-create-hero-image; or wants to interactively design cover imagery in Prisma's 2026 brand (light paper, prism accents, Sora). Produces an editable SVG hero plus a pixel-exact PNG meta image, and includes an interactive mode and a built-in design-review pass.

日本語の概要は準備中です。原文の説明を表示しています。

prisma/web1,1052026年10月10日 更新

Optimize a prisma.io page for search engines and AI answer engines. Use when writing or reviewing blog posts, docs pages, or landing pages for SEO, GEO, AEO, AI citations, AI Overviews, ChatGPT/Perplexity visibility, featured snippets, metadata, or FAQ sections; when refreshing an existing page for freshness or rankings; or when asked why a page isn't ranking or being cited.

日本語の概要は準備中です。原文の説明を表示しています。

prisma/web1,1052026年10月10日 更新

Use when the operator wants to write a blog post, draft a blog article, start a new post for the Prisma blog, or publish to prisma.io/blog.

日本語の概要は準備中です。原文の説明を表示しています。

prisma/web1,1052026年10月10日 更新

Use when adding a new docs section or product area, editing llms.ts / the llms.txt or llms/[...slug] / llms-full.txt routes / get-llm-text / skill.md / .well-known endpoints, or working on the "agent score", "llms.txt", or anything "agent-ready" in the docs and site apps. Explains the invariants the Mintlify agent-readiness audit measures and how to hold them.

日本語の概要は準備中です。原文の説明を表示しています。

prisma/web1,1052026年10月10日 更新

Use when writing, rewriting, or improving technical docs (quickstarts, how-tos, tutorials, concept pages, or API references).

日本語の概要は準備中です。原文の説明を表示しています。

prisma/web1,1052026年10月10日 更新

prisma のスキルをすべて見る

このスキルの問題を報告する