跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
演示文稿

pyramid-audit

Critiques existing writing against the Pyramid Principle and returns ranked findings, not a rewrite. Use when the user has a draft and wants it reviewed or diagnosed. Not for drafting new content in a specific format; use `pyramid-long-form`, `pyramid-presentation`, or `pyramid-short-form` instead.

tyroneross · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add tyroneross/pyramid-principle --skill pyramid-audit

技能说明


name: pyramid-audit description: Critiques existing writing against the Pyramid Principle and returns ranked findings, not a rewrite. Use when the user has a draft and wants it reviewed or diagnosed. Not for drafting new content in a specific format; use pyramid-long-form, pyramid-presentation, or pyramid-short-form instead. user-invocable: false

Pyramid Audit

Purpose

Diagnose user-supplied writing against Barbara Minto's Pyramid Principle. Load pyramid-principle-core for the structural rules. Load pyramid-source-integrity separately only when the user also requests fact or data validation.

This skill audits. It does not rewrite the content, draft a replacement, or explain the entire framework.

Route the Request

Use this skill when the user supplies an email, memo, report, outline, deck storyline, or other professional writing and asks what is structurally wrong, whether it follows the Pyramid Principle, or how to improve its logic.

Hand off when:

  • the user wants new short-form writing → pyramid-short-form
  • the user wants a report or multi-section document drafted → pyramid-long-form
  • the user wants a deck drafted → pyramid-presentation
  • the user wants the framework explained → pyramid-principle-core
  • the user wants claims, sources, calculations, or data checked rather than structure → pyramid-source-integrity

Do not audit without user-supplied content.

Identify the Inputs

Before auditing, the agent identifies:

  1. Content: The exact text, file, or user-provided source to audit.
  2. Reader question: Use the user's question when supplied. Otherwise infer one from the content and label it as inferred.
  3. Medium: Email, memo, report, deck, or fragment. Infer it only when the content makes the medium clear.

The reader question and medium are audit metadata. They are not part of the document and cannot supply a missing Situation, Complication, claim, or evidence.

Run the Audit

The agent performs five checks in order.

1. Governing Thought

  • Locate the single main claim and quote it verbatim, or write Not stated.
  • Check whether it directly answers the reader question.
  • Distinguish the answer from its evidence. When one sentence recommends an action and another sentence gives the reason, the recommendation is the governing thought even if the evidence appears first.
  • Check whether it appears before supporting evidence or detail. If the reader must process support before learning the main claim, flag a buried governing thought.
  • Always run this check. (minto-p5-pyramid-sorting)

2. SCQA Introduction

  • Apply SCQA only when the supplied content contains an opening preamble.
  • A governing thought followed by support is pyramid body, not an implied preamble.
  • A fragment, list, section, or continuation with no opening preamble receives: Not applicable — the supplied content contains no opening preamble.
  • When a preamble exists, rate Situation, Complication, Question, and Answer as Pass, Weak, or Fail using only the supplied text.
  • Do not infer a Situation or Complication from the reader question or assumed shared knowledge. (minto-p22-scqa)

3. Vertical Logic

For each major point, state the question it raises: Why?, How?, or Why do you say that? Check whether the material immediately below answers that question.

Flag a violation only when the supplied text shows a different answer, premature information, or an unsupported hanging assertion. Quote the break and cite minto-p17-vertical-qa.

4. Horizontal Logic

Apply this check only to a peer set of two or more sibling points. A single sentence, statistic, or quantified fact is not a peer set.

For each actual peer set:

  1. Write the candidate parent question.
  2. Name the role every peer performs under that parent, such as reasons, controls, findings, evidence, risks, or steps.
  3. Keep different topics together when they answer the same parent question and perform the same role. Do not classify peers by their nouns alone.
  4. Check mutual exclusivity from the supplied text.
  5. Check collective exhaustiveness only when the supplied content or cited domain evidence establishes the relevant universe. A parent claim such as three controls asserts completeness; it does not prove completeness. For an excerpt with no defined universe, the required result is Exhaustiveness is not verifiable from the supplied content. Do not declare the set complete or invent a missing item.
  6. Check whether the set uses one reasoning mode: deductive or inductive.
  7. Check whether the parent is an inference that summarizes the children rather than a topic label.

Cite minto-p96-mece, minto-p63-deduction-induction, or minto-p5-pyramid-sorting for each violation.

5. Ordering

Assign an order only when the text shows it:

  • deductive — each point leads to the next and ends in a conclusion
  • chronological — explicit time or process sequence
  • structural — explicit parts of a whole or categorical structure
  • comparative — explicit ranking criterion or comparison signal

List position alone does not establish comparative order. When no order is discernible, report that fact and judge whether it materially harms comprehension. Do not invent an ordering rationale. Cite minto-p5-pyramid-sorting for a violation.

Control Audit Claims

Every finding must quote or describe evidence in the supplied content and cite a valid source anchor. A comment with no source anchor is an editorial observation, not a Pyramid Principle violation, and receives no severity label.

Audit the supplied content's structure, not the truth of its factual claims. A scoped factual statement such as all requests received this quarter is not a MECE violation or a request to prove the data independently.

When the user requests both checks, run pyramid-source-integrity and this structural audit independently. Return separate SOURCE INTEGRITY and STRUCTURE sections. A contr

阅读完整来源 ↗