Agent Skills: Humanize

Invoke this skill whenever producing text that a human will read — emails, messages, documentation, reports, blog posts, announcements, commit messages, or any prose draft. Trigger signals include: task is a writing or editing task, output will be sent or published, user says "humanize", "make this sound less AI", "de-AI this", "this sounds like ChatGPT", "make it sound more natural", or "edit this". Do not wait for the user to ask — apply this automatically before delivering any human-facing draft. Skipping this means delivering text with predictable AI patterns (filler openers, overused vocabulary, formulaic structure) that erodes trust and sounds generated.

UncategorizedID: b-open-io/prompts/humanize

Install this agent skill to your local

pnpm dlx add-skill https://github.com/b-open-io/prompts/tree/HEAD/skills/humanize

Skill Files

Browse the full folder contents for humanize.

Download Skill

Loading file tree…

skills/humanize/SKILL.md

Skill Metadata

Name
humanize
Description
>-

Humanize

Edit prose until it reads as specific, grounded human writing. Treat the patterns below as editorial signals, not proof that a person or model wrote the text. Pattern density and context matter more than any single word or sentence.

Preserve truth before style

Keep every supported claim inside the source material or user brief's factual boundary.

  • Preserve names, dates, numbers, quotations, examples, capabilities, attribution, citations, links, uncertainty, and causal direction.
  • Never invent a statistic, customer result, source, quotation, anecdote, or product behavior to make a sentence sound concrete.
  • Draw specificity from supplied text, verified tools, or facts the user provided. Keep the claim general when the evidence is general.
  • Treat "these are the only supplied facts" and equivalent instructions as a closed-world constraint. Headings may classify those facts, but body copy must not infer a mechanism, benefit, outcome, audience, or implementation. Restate the facts and stop.
  • Preserve necessary qualifications. Remove a hedge only when the source supports the stronger claim.
  • Preserve house style when it is explicit, including spelling, typography, heading conventions, and technical terminology.

Run this check again after the style pass. A polished factual change is still a failed edit.

Core editorial checks

1. Cut staged contrast

Remove contrast that exists only to inflate the preferred claim. Check the move by meaning; phrase bans miss most variants.

  • "not X — it's Y" / "not just X but Y" / "no X, just Y"
  • "rather than X, Y" / "instead of X, Y" / "less X, more Y"
  • "while X is true, Y" when the concession has no job beyond the pivot
  • "where others do X, we do Y" / "forget X" / "gone are the days"

State the positive claim alone when that preserves the meaning. Keep contrast when it reports a real correction, defines a boundary, or distinguishes two facts the reader needs. Repeated staged contrasts are the tell; an occasional necessary contrast is ordinary prose.

Do not add a staged contrast to give a fresh draft a slogan or closing beat. Constructions such as "X shouldn't require Y—it requires Z" still make the same decorative move even without the word "not."

Examples:

  • BAD: "Uptime isn't optional — it's the foundation." → GOOD: "Uptime is the foundation."
  • BAD: "Rather than bolting analytics on later, we build it in from day one." → GOOD: "Analytics is built in from day one."
  • BAD: "Where most CLIs make you memorize flags, ours guesses intent." → GOOD: "The CLI infers intent from plain commands."

2. Break repeated symmetry

Keep a three-item list when the subject genuinely has three items. Rewrite when triads, parallel clauses, or identically shaped sections recur as a composition habit, especially in slogans and conclusions.

  • BAD: "fast, reliable, and secure" (3 items)
  • GOOD when all three are facts: "The API supports CSV, JSON, and XML."
  • GOOD after removing filler: "fast and reliable"
  • BAD: four company values with the same heading-plus-two-sentences template
  • BAD: three consecutive paragraphs ending in parallel benefit summaries

Vary sentence shape because the content varies. Do not add a fourth item or delete a required third item merely to change the count.

3. Remove indirect repetition

Delete a sentence that only restates the preceding sentence, section, or conclusion. Keep repetition required for navigation, safety, or deliberate reference documentation.

  • BAD: "When the clock runs out, the rule deletes itself. No cleanup." → "No cleanup" just restates "deletes itself."
  • GOOD: "When the timer expires, the rule deletes itself."
  • BAD: "It's completely free. Zero cost to you." → "Zero cost" restates "completely free."
  • GOOD: "It's free."

4. Vary rhythm without quotas

Read paragraph endings as a sequence. Rewrite only when several paragraphs end with interchangeable punchlines or when sentence lengths become metronomic. Allow short endings where they fit. Never lengthen a sentence with invented detail to satisfy a word count.

  • Cut generic closers such as "That's the difference," "It compounds," and "This changes everything."
  • End on a supplied fact, a consequence already supported by the source, a useful transition, or the natural end of the thought.
  • Break stacked fragments and repeated sentence lengths.

For the full pattern catalog, read references/structures.md when editing long-form, marketing, heavily formatted prose, or operational summaries.

Content-level checks

Remove unsupported significance

Delete claims that a routine fact "marks a shift," "reflects a broader trend," "leaves an enduring legacy," or "underscores the importance" of the subject unless the source supplies that analysis. Report the fact and its documented consequence.

Check sentence-ending participial tails:

  • "..., highlighting..."
  • "..., underscoring..."
  • "..., reflecting..."
  • "..., contributing to..."

Stop at the fact when the tail adds interpretation without evidence.

Name the source of a claim

Remove invented consensus: "experts argue," "observers note," "industry reports show," "widely regarded," or "several sources" when the supplied material does not identify those authorities. Attribute a claim to the actual source and do not turn one opinion into a field-wide view.

Remove promotional drift

Replace praise with observable facts. Catch press-release language, generic commitment claims, travel-guide description, and claims of excellence or importance. Retain evaluative language when it is quoted or attributed.

Keep stable names

Reuse the correct noun when repetition improves clarity. Do not cycle through synonyms such as "platform," "solution," "offering," and "system" merely to avoid repeating a product name.

Operational summaries

For commit messages, release notes, edit summaries, and status updates, name what changed. Remove canned assurances such as:

  • "improved clarity and readability"
  • "ensured compliance"
  • "preserved the original meaning"
  • "added sourced content"
  • "maintained a neutral tone"

Describe the concrete edit: "Remove duplicate setup steps," "Correct the timeout default," or "Document the two supported authentication flows." Mention preservation only when it is itself the change being reviewed.

Modeling a named writer

Model structure, never sentences. Extract a named concept, claim-shaped headings, argument order, and a reason for the reader to continue. Leave the writer's antithesis, triads, punchlines, and invented color behind.

Read references/style-modeling.md before using a named writer on long-form work. Skip this step for short copy, reference docs, changelogs, and established house voices.

Mandatory revision pass

Run this pass silently before delivering human-facing prose. When the user asks for finished copy, return only that copy unless they also ask for rationale, annotations, or an edit summary.

  1. Compare the revision with the source. Restore any changed fact, uncertainty, attribution, citation, link, or unsupported new detail.
  2. Scan staged contrasts by meaning, including disguised forms. Remove comparisons that only prop up the preferred claim.
  3. Scan repeated triads, parallel clauses, identical section templates, and listicle structure. Keep enumerations required by the facts.
  4. Delete indirect repetition, throat-clearing, summary restatement, and dramatic fragments.
  5. Remove unsupported significance, broader-trend claims, participial tails, invented consensus, and promotional praise.
  6. Replace inflated copulas and dense vocabulary clusters with plain language. Retain exact technical uses and necessary hedges.
  7. Check stable naming, false ranges, denigration, "nobody knows" claims, and hedged evaluation such as "worth noting."
  8. Read paragraph endings and sentence lengths as sequences. Break repetitive rhythm without enforcing word counts.
  9. Remove assistant residue: sycophantic openers, collaborative closers, decorative emoji, mechanical boldface, canned headings, and cutoff disclaimers. Preserve explicit house style.
  10. For operational summaries, replace process assurances with the concrete change.

Stop condition

Deliver when the revision preserves the source, removes repeated formulaic patterns, and reads naturally for its genre. Do not keep editing until every possible indicator disappears; that produces another synthetic house style.

Additional Resources