git:20261011.b2d53fa to git:20261024.bebadb9

40 added, 38 removed. Audit A to A.

---

name: humanizing-prose

description: Use when the task itself is to draft, rewrite, edit, or preserve human-facing prose for clarity, attribution, or proportion; do not select merely because other work ends with a response.

when_to_use: "Trigger phrases: rewrite this, make it read human, remove AI slop, no em dashes, tighten this doc, draft the announcement, edit the blog post, polish the PR description, sounds robotic, write it in my style, draft the Slack message, write the Linear or Jira comment, ticket description, write it so a non-engineer understands, write normally."

metadata:

short-description: Draft or edit human-facing prose without AI-slop markers

---

# Humanizing Prose

- Use the available context to identify the speaker, recipient, purpose, and

- what the conversation has settled. Write a message that belongs in that

- exchange. Ask for missing context only when it materially changes the message.

- Follow explicit user requirements for voice and format. When the user asks

- for their own voice, model it on their messages and writing samples available

- in context: sentence length, vocabulary, punctuation, and formatting habits.

- Remove padding, sales language, generic optimism, and irrelevant session history.

+ This skill replaces the length, sentence, and opening defaults of the

+ operator [output style](../../output-styles/megapowers.md). Its attribution

+ and specific-evaluation rules still apply.

- Preserve the requested substance, including every recommendation the user

- asked to share. Keep facts, reasons, caveats, and uncertainty needed to assess

- it. Summarize only within the requested scope; do not omit required detail to

- fit a preferred length or structure. Preserve identifiers, numbers, commands,

- and decisions accurately. Do not rewrite quotations, code, or command names. Never invent

- facts, agreement, ownership, commitments, or experience.

+ Identify the speaker, recipient, purpose, and what is already settled. Ask for

+ missing context only when it changes the substance. Write from an established

+ speaker's perspective, never the assistant's. When the user asks for their own

+ voice, copy the sentence length, vocabulary, punctuation, capitalization, and

+ formatting of their human-facing writing, not their prompts to the agent. When

+ editing, keep sentences that work and change only defects.

- Keep proposals, questions, decisions, and commitments distinct. Use the author's

- perspective when established, without turning a proposed assignment into a

- promise. Choose paragraphs, lists, and headings for the reader's task; do not

- copy the assistant's analysis structure by default. Leave already-effective

- prose alone.

+ Preserve the requested substance, including every recommendation the user

+ asked to share, with the facts, reasons, caveats, and open questions needed to

+ judge it. Do not drop required detail to fit a length. Keep quotations, code,

+ commands, identifiers, numbers, and decisions exact. State a verification

+ limit inside the sentence, not as an appended disclaimer. Never invent facts,

+ test runs, agreement, ownership, commitments, or experience.

- For examples of discussion and review structure, see

- [workplace comments](references/workplace-comments.md).

+ Keep proposals, questions, decisions, and commitments distinct. Do not turn an

+ offer into a promise. Lead with the point in reader order, not the order of

+ investigation. Choose paragraphs, lists, and headings for the reader's task.

## Machine-prose markers

- Apply the plugin [output style](../../output-styles/megapowers.md) rules for

- attribution and specific evaluation; load it if it is not active. Do not use

- em dashes or spaced hyphens as em dashes, even when the surrounding archive

- uses them. Remove triads of adjectives, "not X but Y" framing, rhetorical

- questions, "it's worth noting", "in today's landscape", summary headers that

- restate the body, and generic closing offers or recaps. Vary sentence length

- instead of stacking parallel clauses. Keep one register across the piece.

- Replace unmeasured intensifiers with a number, bounded scope, or source.

- Collapse stacked hedges to one confidence level while preserving material

- uncertainty.

+ The quoted phrases are examples. Cut the move, not only the phrase.

- For a status update or review summary, give the current verdict, material

- impact, and minimum evidence in plain paragraphs without bold labels; do not

- repeat findings already posted inline. For a discussion reply, include the requested

- recommendations, useful reasons, and unresolved questions. Check that the

- recipient can understand what the message adds and any response needed.

- Do not publish routine progress narration, session handoffs, repeated findings,

- or test transcripts as tracker, issue, or PR comments. When publication is

- authorized, reply in the existing thread when possible.

+ - No em dashes or spaced hyphens as dashes, even if the source uses them.

+ - No `**Label:**` lead-ins unless a template requires them, no headings on

+ short messages, or summary headers that restate the body.

+ - No "not X but Y" framing, rhetorical questions, or announcements of

+ importance ("it's worth noting", "crucially").

+ - No padding, sales language, generic optimism, session history, opening

+ praise, or closing offers and recaps.

+ - Vary sentence length instead of stacking parallel clauses or adjective

+ triads. Keep one register. Use the same noun for the same referent.

+ - Collapse stacked hedges to one confidence level while preserving material

+ uncertainty.

+ ## Workplace threads and reviews

+ For a status update or review summary, give the verdict, material impact, and

+ minimum evidence in plain paragraphs without bold labels. Do not repeat

+ findings already posted inline. In a discussion reply, open with your answer,

+ stance, or question, not a restatement of the other message. Do not post

+ progress narration, session handoffs, repeated findings, or test transcripts

+ as comments. When publication is authorized, reply in the existing thread.

+ Examples: [workplace comments](references/workplace-comments.md).