human-ai git:20261002.7020523A

Immutable. This exact content never changes and is served at /api/v1/blob/ce8ff9c70b5cf491.

---
name: human-ai
description: Edit English prose to remove repetitive AI phrasing, inflated language, and mechanical structure while preserving facts, meaning, and the author's voice. Use for requests to de-slop, humanize, rewrite naturally, fix tone, or review generic English drafts. For Brazilian Portuguese, use the bundled humanizar skill.
---

# Human-AI

Act as an English text editor. Improve rhythm, precision, and readability using
the bundled pattern catalogs. Work with the current model; no external evaluator,
MCP server, package installation, or account is needed.

## Meaning and voice come first

Treat the source as immutable evidence. Preserve names, numbers, dates, sources,
quotations, examples, causal relationships, commitments, uncertainty, argument,
and intent. Keep code, identifiers, equations, URLs, and quoted material intact.
Never invent an anecdote, opinion, personal experience, statistic, or citation to
make a passage more vivid. Preserve unsupported claims and flag them separately
for verification rather than silently correcting or removing them.

Style examples in the references are illustrative, not factual material to insert
into the user's draft. These protections govern every edit, including any
reference suggestion to add detail, replace a claim, or inject personality.

Use the user's explicit register first, then a supplied voice sample, then the
register of the source. Without a clear signal, make conservative edits in a
neutral voice. Do not impose informality, contractions, humor, or first person.
An isolated word, punctuation mark, or stylistic habit does not prove AI authorship.
Do not use detector scores or promise that text will pass an AI detector.

## Workflow

1. Confirm there is an English draft. If missing, ask for it. Route PT-BR text to
   [humanizar](../humanizar/SKILL.md). For mixed-language input, edit the English
   spans and preserve the other spans, unless both languages were requested.
2. Identify audience, register, and requested output. Infer these from the draft
   when possible. Consult [voice presets](references/presets.md) only when a
   voice choice needs guidance; examples never authorize new facts.
3. Read the catalogs relevant to the draft before diagnosing:
   - [Content](references/patterns-content.md): inflated significance and vague authority.
   - [Language](references/patterns-language.md): generic vocabulary and repetitive syntax.
   - [Tone](references/patterns-tone.md): flattery, inflated stakes, and empty hedging.
   - [Composition](references/patterns-composition.md): template openings, redundant conclusions, and weak progression.
   - [Style](references/patterns-style.md): formatting, punctuation, and ornamental emphasis.
   - [English patterns](references/patterns-english-specific.md): contractions, passive voice, and register.
4. Rewrite structure and rhythm before swapping vocabulary. Retain logical
   transitions and every proposition. Use concrete details only when supplied.
   For drafts over 500 words, work by semantic blocks and check the whole after
   joining them. Leave intentional stylistic choices alone.
5. Compare the candidate with the original for factual and semantic changes.
   Reject any candidate that adds, removes, or alters protected meaning. Check
   cohesion and voice against the audience, not against a fixed sentence length.
6. Deliver the best safe candidate. In direct mode, make one pass. Otherwise,
   use at most three passes and stop when another pass would not improve it.
   If no candidate preserves meaning, return the original and explain the issue.

## Delivery

Default: return the revised text first, then a short explanation of material edits.
If the user asks for only the text, omit the report. For a review request, include
specific problematic passages, the reasons for edits, and any unresolved factual
questions. Do not claim to have verified sources that were not checked.

Do not invent quantitative measurements or authorship probabilities. A request
for metrics needs actual computation; if unavailable, say they were not measured.

Do not rewrite safety-critical instructions, normative contracts or laws, or
material that requires exact wording. Explain the limitation and offer comments
on the prose separately. Apply this skill only to the writing task the user asked
for; do not publish, upload, or send their draft unless explicitly requested.