/slashskills
back

humanize

Edit or audit prose for generic AI patterns while preserving voice, meaning, uncertainty, and format. Use when writing sounds synthetic, repetitive, over-polished, or unlike its author.

Category
workflow
Package
humanize/SKILL.md
License
MIT
Author
@tushaarmehtaa
Tags
writingeditingvoiceai-slopprosetone

Install

Swipe for more runtimes.

Codex

Skills directory: ~/.agents/skills

Install globally

$npx skills add tushaarmehtaa/tushar-skills --skill humanize -g -a codex -y

Invoke

$humanize or /skills

You can also describe the task naturally; runtimes may select the skill from its description.

Required access

files you provide
Setup guide →Official docs ↗

Claude app

This workflow can run in chat using the files and context you provide. Download its complete ZIP, then upload it from Claude's Skills settings.

↓ download ZIPUpload guide →

ChatGPT Skills

This workflow is suitable for ChatGPT Skills. ChatGPT does not document the same upload archive format as Claude, so follow its uploader instead of reusing the Claude ZIP.

ChatGPT upload guide →

Instructions

Source: SKILL.md

Humanize

Make prose sound like its author writing clearly, not like a different fashionable voice. Remove generic model habits only when the text supplies evidence for the edit. Preserve facts, calibrated uncertainty, technical terms, personality, and the register required by the format.

Keep the boundary clear

Use this skill for voice preservation and synthetic-writing patterns.

  • Use landing-copy, cold-outreach, readme, or ui-copy when the primary job is creating or correcting that format's substance.
  • Use remove-ai-slop for rendered interface patterns and cross-route design convergence.
  • Use an article-development workflow when the argument, reporting, or structure needs major editorial work before a prose pass.

A request to “humanize,” “de-slop,” preserve voice, or detect AI-like patterns belongs here. A generic “make this better” should stay with the format-specific skill unless the evidence shows voice is the main problem.

Choose the mode

  • Edit — return revised prose while preserving the author's position and factual meaning.
  • Detect — identify checkable patterns and explain their effect without rewriting.
  • Calibrate — derive a compact voice profile from supplied samples for a later or included edit.

Infer the mode and state it briefly. Do not claim that a detector can determine who or what wrote the text, and do not output an “AI percentage.”

Establish the writing contract

Read the full draft before changing sentences. Determine:

  • audience, format, and intended reader action or belief;
  • point of view, vocabulary, cadence, humor, formality, and recurring intentional habits;
  • statements that carry facts, uncertainty, legal scope, attribution, or personal experience;
  • passages that already sound specific and alive.

Infer the voice from a substantial mixed-quality draft. If the text is almost entirely synthetic and the user explicitly wants their personal voice, ask for one representative sample in a single concise question. If they only want clearer prose, proceed in an appropriate neutral register and say that no personal voice sample was available.

Never ask for a sample merely to avoid editing.

Diagnose before rewriting

Read references/patterns.md for the evidence tiers. Separate findings into:

  • Substantive defects: unsupported claims, invented evidence, semantic ambiguity, missing attribution, or lost qualification.
  • High-confidence defaults: language that adds no information or disguises the actual claim.
  • Contextual patterns: ordinary devices that become generic only through repetition, mismatch, or dominance.

Do not treat punctuation, vocabulary, sentence length, active voice, fragments, jargon, lists, or rhetorical questions as proof in isolation.

Edit proportionally

  1. Preserve strong human passages and exact useful details.
  2. Remove empty setup, duplicated meaning, unsupported importance, and placeholder scaffolding.
  3. Replace vague abstraction with a concrete statement only when the draft contains that information.
  4. Vary sentence shape only when the existing cadence is monotonous; do not manufacture decorative variation.
  5. Keep passive voice when the actor is unknown, irrelevant, deliberately de-emphasized, or conventional for the format.
  6. Protect hedges that express real uncertainty or scope.
  7. Read references/formats.md and adapt the pass to the deliverable.

Never invent a statistic, example, quotation, source, opinion, anecdote, or product capability to make a sentence more vivid.

Resolve consequential edits interactively

Make ordinary style edits without asking. Pause or mark Needs you when a proposed change would alter the author's stance, remove a real qualification, choose between conflicting voice samples, or require a fact the draft does not supply.

When the missing choice blocks the whole edit, ask one question with a recommended interpretation. Otherwise complete the safe edit and isolate the unresolved passages.

Output

Adapt the output to the size and mode.

For edit mode:

  1. Return the complete edited draft without inline commentary.
  2. Add a short What changed section grouped by material pattern or structural decision.
  3. Add Preserved when naming deliberate voice choices helps the user trust the edit.
  4. Add Needs you only for unresolved facts, claims, or voice decisions.

For a one-line or very short edit, return the edit and one sentence of rationale; do not force four empty sections.

For detect mode, list the pattern, exact excerpt, evidence tier, consequence, and repair direction. Also identify specific passages that already work. Do not rewrite unless requested.

Verify

Read references/evaluation.md before returning substantial work. Confirm that:

  • the author would recognize the position and voice;
  • factual meaning and calibrated uncertainty survived;
  • the amount changed matches the evidence;
  • the revision does not install uniform staccato, corporate polish, or a terse personal house style;
  • the result works in its actual format when read aloud or exercised in context.

Bundled references

3 files · 147 lines

references/evaluation.md

source ↗

Humanize evaluation and sources

Use this checklist after a substantial edit. Fix failures before returning the work or place consequential unknowns in Needs you.

Voice and meaning

  • Can the core point be stated the same way before and after?
  • Did distinctive vocabulary, cadence, humor, uncertainty, and personal detail survive?
  • Are strong original sentences preserved rather than polished for consistency?
  • Did any edit change attribution, causality, chronology, comparison, scope, or confidence?
  • Is the amount changed proportional to the diagnosed pattern?

Nothing invented

  • Are there no new facts, statistics, quotations, examples, sources, product abilities, or opinions?
  • Are unsupported claims identified rather than quietly strengthened?
  • Are factual defects separated from stylistic findings?

Over-correction

  • Does the result avoid uniform sentence length, stacked punch lines, and repeated “plain” sentence shapes?
  • Did the edit preserve useful jargon, passive constructions, repetition, and hedges where the format needs them?
  • Does the prose sound like the source author rather than the editor's house style?
  • For short copy, was output kept short instead of forcing a large editorial report?

Format test

  • Does the piece still accomplish its actual job?
  • Is its hierarchy scannable where readers scan and continuous where they need an argument?
  • When the format is spoken, does it work aloud?
  • When the format is operational, are instructions, labels, and consequences unambiguous?

Sources and provenance

This package reconstructs Tushar Mehta's earlier humanize draft, which credited the MIT-licensed remove-ai-slop and no-ai-slop skills plus the Slashskills design-audit verdict system. This version retains the voice fingerprint, evidence tiers, format register, and over-correction guard while removing universal bans and rigid output requirements.

references/formats.md

source ↗

Format register

Read only the section matching the deliverable. Format requirements outrank a generic preference for conversational prose.

Social posts and threads

Protect brevity, fragments, lowercase, direct opinion, and native informality when they belong to the author. Remove unearned hooks, engagement-bait questions, templated lesson lists, and one-line staccato used through the entire post.

Newsletters and personal essays

Protect digressions, uneven paragraphs, jokes, and recurring personal habits that create the relationship with the reader. Remove generic greetings, unnecessary roadmaps, repetitive recaps, and endings that inflate a specific observation into universal wisdom.

Reported or analytical essays

Protect the argument's necessary setup, sourced uncertainty, long sentences doing real logical work, and terminology the subject requires. Remove duplicated conclusions, vague authority, ornamental significance, and metaphors that replace the mechanism.

Landing and marketing copy

Preserve product nouns, concise declaratives, and repeated terminology. Test whether a claim could be pasted onto several unrelated competitors. Replace generic benefit language only with capabilities, evidence, or outcomes supported by the source material.

Documentation and READMEs

Prioritize a working outcome, exact terminology, scannable structure, code or tables where they reduce ambiguity, and explicit responsibility. Passive voice may be correct when the actor is the system or irrelevant. Do not inject marketing cadence or personality into procedural reference material.

Interface copy

Match the user's state and the system's actual knowledge. Preserve concise controls and conventional labels when they are clearest. Never make errors, destructive actions, payment consequences, privacy choices, or permissions playful merely to sound human.

Cold outreach

Protect directness and a specific observation that proves attention. Remove reusable flattery, sender autobiography before relevance, and asks buried under politeness. Do not invent personalization.

Video and spoken scripts

Read the revision aloud. Protect useful repetition, breathing room, fragments, and conversational transitions. Remove written constructions that trip the tongue or retention hooks that promise more than the script delivers.

Legal, medical, financial, and policy text

Do not trade accuracy, mandated language, provenance, or calibrated uncertainty for warmth. Flag ambiguous or inaccessible language, but require appropriate subject-matter review before changing consequential meaning.

references/patterns.md

source ↗

Writing-pattern evidence

Use this catalog to diagnose, not to run mechanical find-and-replace. The same construction can be useful in one passage and generic in another.

Evidence tiers

Substantive defect

Report separately from style:

  • unsupported or unverifiable facts, comparisons, testimonials, or statistics;
  • attribution that hides the source behind “experts,” “research,” or “people say”;
  • a rewrite that removes meaningful uncertainty, legal scope, or technical precision;
  • placeholders presented as finished content;
  • product terminology that contradicts observed behavior.

Do not repair a missing fact by inventing one.

High-confidence default

One occurrence may justify an edit when the construction contributes no information in context:

  • throat-clearing that only announces a point;
  • importance language that tells the reader how consequential a fact is instead of showing why;
  • a question answered immediately for theatrical effect;
  • a closing aphorism unrelated to the concrete conclusion;
  • adjacent passages that make the same claim without new detail;
  • generic transitions left from drafting scaffolds.

Even here, preserve a construction when it carries a documented voice function, real contrast, pacing, or necessary orientation.

Strong presumption

Require contextual justification before keeping repeated patterns such as:

  • manufactured binary reversals or clipped negative lists;
  • padded triplets whose third item exists only for cadence;
  • benefit language that obscures the capability or mechanism;
  • inflated verbs and nominalizations where a plain verb is clearer;
  • empty superlatives and adjectives without a measurable referent;
  • synonym cycling that makes one concept look like several;
  • audience flattery that avoids naming the real reader;
  • decorative formatting that does not improve navigation or emphasis.

Contextual signal

Flag only when repeated, mismatched to the format, or costly to comprehension:

  • em dashes, fragments, rhetorical questions, contractions, passive voice;
  • jargon, long sentences, short paragraphs, bullets, bold text, emoji;
  • first or second person, humor, asides, metaphors, parentheticals;
  • sentences beginning with conjunctions;
  • repeated words used intentionally for clarity.

Tests for a proposed edit

Ask:

  1. Does deleting or changing this lose a claim, relationship, voice cue, or useful pace?
  2. Is the problem the device itself, or its repetition and placement?
  3. Does the replacement use information actually present in the draft?
  4. Would the author's strong untouched sentences plausibly contain the replacement?
  5. Does the format reward this construction—for example, repetition in audio or exact terminology in documentation?

If the evidence is weak, leave the line alone.

edit source on GitHub ↗