首页/技术写作/reviewing-technical-prose
R

reviewing-technical-prose

by @riekeltv
4.6(120)

此技能用于审查、重写或编辑他人的技术文本,生成审查报告,并在文档交付前做最终检查。它内建严重级别映射、发现格式、不应标记的清单和交付检查清单,帮助用户系统化地识别文档缺陷,保证内容准确、结构清晰、语言自然。适用于任何需要保证技术文档质量的场景。

technical-writingdocumentationeditingreviewproseGitHub
安装方式
npx skills add https://github.com/riekelt/technical-writer --skill reviewing-technical-prose
compare_arrows

Before / After 效果对比

1 组
使用前

在没有这个技能之前,用户手动逐段审查技术文档,没有统一的严重度标准和检查清单,容易遗漏关键问题(如不支持的声明、机器腔调),一份 20 页文档通常需要 4 小时以上,且一次审查后仍可能残留缺陷。

使用后

使用此技能后,审查流程被结构化:按照严重度映射、发现格式和不应标记清单快速定位问题,一份 20 页文档的平均审查时间降至 30 分钟,缺陷发现更全面,文档交付质量显著提升。

SKILL.md

Reviewing technical prose

REQUIRED BACKGROUND: the technical-writing skill, including references/style.md (banned constructions) and references/truth.md (claim rules).

Overview

Editing is diagnosis. Every edit names the concrete defect it fixes; the smallest edit that fixes it wins. Rewriting natural or approved language without a named defect is itself a defect.

Review the full document, never a summary of it: a reviewer working from a digest invents missing-section findings.

When to invoke, and not

Invoke when reviewing or rewriting someone else's technical text, when producing a findings report on a document, and as the final pass before delivering anything you wrote yourself. Do NOT invoke for code review (only the prose in it), and do not use a review pass to relitigate settled decisions, expand scope, or restyle a document onto your own preferences.

Severity mapping

  • BLOCKER: hard-rule violations (banned dashes, changelog sections, delivery history in prose) and any claim the cited source does not support or that traces to nothing.
  • WARNING: banned constructions, structural defects (question headings, buried conclusions, rearrangeable paragraphs), and claims supported only loosely.
  • OBS: ambiguity, missing polish, and anything an attentive author would likely catch.

Rewriting someone else's text

One requirement above all rules: the content stays identical. Work in this order:

  1. Read the source. Mark every banned construction and word-choice violation.

  2. Write a version. Read it aloud.

  3. Ask three control questions:

    • Which sentence still sounds like a language model?
    • Did the rewrite add or drop a fact, number, date, name, source, or claim?
    • Does the rewrite recreate a removed pattern in a new rhetorical form?

    Repairs breed their own tells, so re-scan the rewritten paragraph as if it were source text.

  4. Repair what the control questions surface. An addition and a loss both count as errors, even when the text reads better for it.

Never add a fact to finish a sentence: ask the author or pick a simpler sentence. If a sentence stays wooden after two attempts, rewrite the whole paragraph around its main point. Change only running prose: code blocks, frontmatter, table data, and link targets stay as they are.

  • The rewrite keeps the document's language. A Dutch document comes back in Dutch, with the structural and truth rules applied as always; the English-specific vocabulary checks are replaced by that language's own list where one exists.
  • When the author is reachable, the diff is the proposal. Show the rewrite before overwriting their text. Unattended in a repository, the commit is the proposal and review does the same job.

When the source text asserts nothing recoverable (setups, glosses, meaning-sentences), the correct rewrite is deletion plus a marked gap (**[input wanted: <the claim the sentence should make>]**, the claim-level sibling of the core skill's **[source wanted: ...]**) where a real claim should stand. A rewrite full of marked gaps is the intended outcome for claim-free source text, not a failure.

Findings to leave alone

The banned-constructions list helps recognize machine text; it proves nothing by itself, and every pattern also occurs in good human writing. Flag only when several signs coincide in the same paragraph.

  • Polished grammar and consistent formatting: many writers are professionals, or edited.
  • One repeated sentence opening: repetition can be rhythm. Fix only when it adds nothing.
  • One short sentence for emphasis: only a row of fragments is a problem.
  • Formal words in general: the list names specific words; not every formal word needs simplifying.
  • A factual contrast: "the pipeline sets the tag, the developer does not" is a statement, not antithesis.
  • A serious alternative the reader would genuinely weigh: it belongs in the document.
  • Qualifiers that bound something: scope, assumptions, and safety or legal notes stay. Only the stacking goes.
  • A reference to the previous situation in a migration doc or release note: there the change IS the subject.
  • A quoted word: never rewrite inside a citation, a title, or an example that discusses the word.
  • Metaphor that explains: only the decoration around it goes.
  • Deliberate awkwardness: a clear fragment or lopsided sentence is not a defect merely because it could be polished.

Review reports

  • Verdict in the first line, findings ordered most severe first.
  • Severity vocabulary defined in the report that uses it: BLOCKER (cannot proceed), WARNING (likely rework; fix but not blocking), OBS (worth noting).
  • Every finding: one sentence stating the defect, a location (file:line, section, task ID), a proposed fix, and a confidence level.
  • A named empty case ("FINDINGS: none") so a silent reviewer and a clean result cannot be confused.
  • A document claim contradicted by the code is a defect in the document; report it with the conflicting source. An unrelated bug you noticed in passing is not your finding.
  • Re-reviews list only what remains.
  • Acknowledge what is good; a review that only objects is not calibrated.

Delivery checklist

Before any document goes out:

  • Matched to the existing documents in the same directory, and added to the README.md index if one exists
  • No em dashes, en dashes, or --; no changelog section or "last updated" field; no ticket keys, phases, or SHAs in prose
  • No banned constructions (the technical-writing skill's references/style.md); check at least the summary and the closing paragraphs
  • Headings name the content: no question forms, no "What X means", no heading repeated in its first sentence
  • Conclusion first at document, chapter, and paragraph level
  • Procedures: one action per sentence, with an actor
  • Where the document proposes or decides: every non-trivial choice has its reasoning and a serious alternative recorded, and the costs are in, not only the benefits
  • Fact and proposal are distinguishable; estimates labeled; unknowns explicitly unknown
  • Every claim with a number traces to a source named in the document
  • References checked by following them: chapter numbers, relative links, and file names actually exist (they break silently on every restructure)
  • Terms consistent and defined at first use; one term per concept
  • Read aloud: at least the summary; no row of fragments, no three sentences with the same opening
  • Remove-the-name test on the opening and summary: with the product name deleted, a stranger can still tell what the text is about
  • Reads cold, reads whole: natural to someone who never saw the conversation or feedback rounds that produced it; one voice throughout, no patchwork seams, no sentence explaining why the document was written this way
  • On a rewrite: no fact, number, date, or source added or lost

用户评价 (0)

发表评价

效果
易用性
文档
兼容性

暂无评价

统计数据

安装量17.3K
评分4.6 / 5.0
版本
更新日期2026年10月23日
对比案例1 组

用户评分

4.6(120)
5
37%
4
43%
3
13%
2
5%
1
2%

为此 Skill 评分

0.0

兼容平台

🤖claude-code

时间线

创建2026年8月30日
最后更新2026年10月23日
🎁 Agent 知识卡片
调研问卷