Skip to main content
skill CX Codex

human-ai · fabricioctelles, deslop-writing — Claude Code skill: what it does, how to install

human-ai reworks English prose to remove canned AI wording and stiff structure while keeping the source facts, meaning, and authorial voice.

Apache-2.0 ★ 141 updated 2026-10-29

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add fabricioctelles/skills --skill human-ai -a claude-code
Open on GitHub

What human-ai does

The skill edits English drafts, notes, or transcripts when the wording feels generic, inflated, or mechanically arranged. It protects names, numbers, dates, quotations, causal links, uncertainty, deadlines, and commitments instead of smoothing them away. It can also compose a draft from the author’s own material.

Its workflow checks the source claims before and after rewriting, then uses pattern catalogs for content, language, tone, composition, and style. When an essential detail is missing, it can ask up to three targeted questions. Brazilian Portuguese is routed to the bundled humanizar skill. No external evaluator, MCP server, package install, or account is needed; it runs with the current model. That makes the original material and model output the practical limits on the edit.

What it helps with

  • remove canned AI phrasing from English prose
  • keep source facts intact during a rewrite
  • turn notes or a transcript into an English draft
  • review a passage for stiff rhythm and inflated wording

Example prompts for the agent

  • Rewrite this English draft in a more natural voice, keeping every fact and the author’s intent.
  • Review this copy for canned AI phrasing, then revise it without adding claims.
  • Draft English prose from these notes and ask questions only where a material fact is missing.

How to install human-ai

1Skills CLI (Vercel)

npx skills add fabricioctelles/skills --skill human-ai -a claude-code

2Manually: put the files in place

~/.claude/skills/human-ai/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/human-ai/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about human-ai

How do I install human-ai in Claude Code?

Run npx skills add fabricioctelles/skills --skill human-ai -a claude-code, then check ~/.claude/skills/human-ai/SKILL.md.

What does human-ai do?

It edits English prose to reduce repetitive AI wording and rigid structure while preserving the source meaning, facts, and authorial voice.

Does human-ai require an MCP server or package installation?

No. It runs with the current model and does not need an external evaluator, MCP server, package installation, or account.

Source text

SKILL.md · Apache-2.0
---
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. Also compose from the author's English notes or transcripts and ask targeted questions when real details are missing. 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, including actors, temporal state, deadlines, scope, and exceptions.
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.
Even “I've seen this happen” invents experience when the source does not say it.
Do not add an apology, promise, benefit, or eligibility condition to sound warm.
Only explicit authorization permits substantive cuts or changes. A failed factual
or semantic lock invalidates the candidate before style is assessed.

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.

## Notes and targeted questions

Notes, transcripts, and rough personal drafts are sufficient input. Use them as
the main material: preserve claim order, causal reasoning, examples, expressions,
and sentences where possible. An accompanying AI draft is a checklist of possible
points, not a substitute for the author's argument. Map claims to their source.
Facts found only in the AI draft are unverified; do not merge or delete them
silently. Ask about a material conflict unless source precedence is already clear.

Interview when requested, or when an essential missing fact prevents useful,
faithful writing and the user can answer. Ask one round of up to three questions,
each tied to a real passage or gap. Seek the author's example, number, event, or
reasoning in their own words. Infer tone and audience when possible. Complete
operational drafts need no interview. In direct mode or unattended pipelines,
edit conservatively and identify gaps instead of asking optional questions.

Answers become approved factual material with provenance; preserve the original
artifacts. A skipped or unanswered question does not authorize invention or cuts.
Keep supported general wording, flag the missing detail, or pause if an essential
answer is required. Ask again only when a new contradiction blocks the task.

## Workflow

1. Confirm there is English material: a draft, notes, or transcript. If absent,
   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. Inventory the source claims, actors, degree of certainty, deadlines, conditions,
   calls to action, and provenance. Keep this compact and internal unless requested.
   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. Audit both directions against the original and approved user additions: every
   protected proposition survives at the same strength, and every new assertion
   has support. Possibility must not become a promise, proposal must not become
   commitment, acknowledgment must not become apology, and an offer must not gain
   a condition. Preserve who does what. 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.
Keep gaps outside publishable prose unless a working draft was requested. Offer
a shorter option only when requested or substantial redundancy makes it useful;
explain substantive cuts before applying them. Do not append a mandatory footer.

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.

## Provenance

The inventory, bidirectional audit, and notes/interview guidance are adapted from
[phuryn/work-humanizer](https://github.com/phuryn/work-humanizer), reviewed at
commit `0a9065791d021f3dda695f3887cb2ae4c0eacbcf`. Its evaluations do not establish
performance for this skill. The workflow uses the current model and bundled
references; no source download, external service, or detector is required.

© Fabricio Telles · Apache-2.0 · fabricioctelles/skills

Similar skills

All skills for Writing and docs →

human-ai by other authors comes first.

human-ai https://ft.ia.br The human-ai skill edits English prose to remove AI-like phrasing and stiff structure while keeping the writer’s meaning, facts, and voice intact. CX Codex +1 ★ 141repo stars 02_english_prose centminmod 02_english_prose rewrites an English paragraph to remove every comma. You can restructure sentences and use other punctuation while preserving the meaning. GP ChatGPT ★ 2,744repo stars uk-business-english jezweb uk-business-english — a skill that gives your agent British business English: EN-GB spelling, understated tone, the right greetings and sign-offs. Useful when a British audience reads the text. CX Codex +1 1,100installs stop-slop Hardik Pandya (https://hvpandya.com) The stop-slop skill removes predictable AI writing patterns from prose. Use it when drafting, editing, or reviewing text for formulaic phrasing. CX Codex +1 20,300installs sepia-refactor Nanako0129 sepia-refactor — a Claude Code skill for minimal in-place prose revision. Useful when you need controlled text edits without rewriting or handing tool permissions to the agent. CX Codex +1 1,400installs better-writing forjd The better-writing skill rewrites, drafts, and reviews prose for its audience, purpose, and format. Use it for emails, docs, UI copy, marketing text, and posts. CX Codex +1 624installs creative-writing-muse haowjy The creative-writing-muse skill turns the agent into a co-author: discusses the plot, writes drafts in your voice, critiques and revises the text, keeps project memory in one conversation. CX Codex +1 568installs

From the same set46

The whole set →
ultimate-design-system-master ultimate-design-system-master guides users through an 18-step briefing questionnaire to produce design systems, brand assets, and UI code across 10 specialized generators. CX Codex +1 ★ 141repo stars substack-expert substack-expert — a skill for shaping Substack newsletters, from post structure and SEO fields to reader engagement and free-to-paid conversion. CX Codex +1 ★ 141repo stars pier-cloud A skill for calling the Pier Cloud (Lighthouse) API: JWT auth, contexts, workspaces and FinOps view data through ready-made Python scripts. CX Codex +1 ★ 141repo stars geo-optimization geo-optimization optimizes site content for LLM citations and answer engines like ChatGPT, Perplexity, Claude, and Google AI Overview. CX Codex +1 ★ 141repo stars design-md-validator design-md-validator is a Claude Code skill that checks DESIGN.md against Google's specification with the @google/design.md CLI linter. It flags broken token references and computes WCAG contrast ratios. CX Codex +1 ★ 141repo stars humanizar humanizar is a Claude Code skill that drafts and edits Brazilian Portuguese text with natural cadence while preserving the author's facts and voice. CX Codex +1 ★ 141repo stars agent-plugin-eval agent-plugin-eval audits Agent Plugin repositories against the specification, scores them, and compares two targets. Use it for release reviews and conformance checks. CX Codex +1 ★ 141repo stars loop-architect loop-architect is a Claude Code skill that guides iterative workflow design, from goal refinement and verification criteria to reviewer selection and termination guards. CX Codex +1 ★ 141repo stars

Project sponsors

Companies that keep this analytics open