gbalabanli/humanize-text

Humanize Text

Use this skill to turn stiff or obviously AI-shaped writing into clean human prose without drifting from the source.

Workflow

1. Classify the rewrite target before changing anything.

• Identify the text type: marketing, email, essay, article, technical, social, or general.
• Identify the requested rewrite intensity: light, standard, or strong.
• Identify hard constraints: tone, length, audience, format, required phrases, citations, links, and terminology.
• If the user supplied detector output, treat it as a hint, not proof.

2. Preserve what must not move.

• Keep facts, numbers, dates, names, claims, links, and quoted language unless the user explicitly asked to rewrite them.
• Preserve structure when structure is functional, such as bullets, headings, steps, or compliance wording.
• Keep domain terminology when simplification would reduce precision.
• If the source is already personal or opinionated, preserve that voice instead of flattening it.

3. Remove high-risk AI patterns first.

• Cut inflated significance claims, vague authority phrases, and generic scene-setting.
• Replace padded adjective stacks with one concrete claim.
• Remove filler transitions and summary lines that say little.
• Collapse repetitive sentence shapes and over-explained connectors.
• Replace generic abstractions with concrete nouns, verbs, and specifics already present in the text.
• Use ai-patterns.md as the checklist.

4. Rebuild the prose in the right mode.

• Use rewrite-modes.md to match the genre.
• light: tighten wording, vary cadence, remove obvious tells, keep structure nearly intact.
• standard: rewrite most sentences for natural flow while preserving structure and meaning.
• strong: reorganize locally, compress repetition, and sharpen voice, but still preserve claims and scope.
• Do not "humanize" by adding slang, fake mistakes, fake anecdotes, or unsupported opinions.

5. Run a credibility pass.

• Read the rewrite against the source and confirm that no fact, qualifier, or caveat disappeared.
• Remove any new claims, emotional framing, or implied authority not grounded in the source.
• Check whether the result sounds like a competent person wrote it, not a detector-avoidance gimmick.
• If the rewrite still sounds synthetic, make it more specific, shorter, and less symmetrical.

6. Deliver the result in the smallest useful format.

• By default, return the rewritten text first.
• Add a brief change summary only when it helps the user understand the tradeoffs.
• If the user asked for detector-oriented help, call out the patterns removed instead of promising a score.

7. If scoring is requested, run a retry loop.

• Score the original text first.
• Humanize and score again.
• If score does not improve enough, diagnose likely causes from detector feedback (e.g., monotone cadence, triplet phrasing, corporate tone) and retry humanization.
• Stop after a bounded number of retries and return the best-scoring rewrite with a short retry log.

Rewrite Rules

• Prefer specificity over polish.
• Prefer direct verbs over abstract nouns.
• Prefer natural sentence length variation over perfectly even rhythm.
• Keep emphasis rare and earned.
• Keep transitions minimal; move directly to the next point when possible.
• Keep personality appropriate to the genre. Technical docs and formal emails should not sound like personal essays.
• Do not manufacture "human" traits such as typos, emoji clutter, forced humor, or fake vulnerability.
• Do not claim detector safety or guarantee a text will pass any AI checker.

Output Rules

• If the user asked only for a rewrite, provide the rewrite without a long preamble.
• If there are preservation risks, state them briefly before rewriting.
• If the source is too short for meaningful humanization, say so and apply only a light pass.
• If the source is formal, optimize for credibility and clarity, not personality.
• If the source is personal or editorial, allow more rhythm and voice, but keep it grounded in the original meaning.

References

• ai-patterns.md
• rewrite-modes.md

Local Tools

• Launch the local detector UI with python scripts/serve_ui.py.
• Use the app in ui/ to paste full text, fetch live AiDetector.com, QuillBot, and Scribbr scores, and run Humanize + Recheck for automatic retry-based humanization with before/after scoring.
• Treat external detector scores as advisory signals. Do not represent them as proof.