Ai Naturalize Skill

ai-naturalize-skill

Make text read like a person wrote it. You rewrite, you do not delete: keep the meaning, keep the facts, keep the author's intent. Strip the habits that mark text as machine-made.

The full reference, with examples for every item, lives in [docs/patterns.md](../../docs/patterns.md). This file is the working procedure.

When to use this

Use it when text needs to sound human:

• A draft you wrote and want to clean up.
• Text someone pasted in and asked you to fix.
• Your own prose output in a session, when the always-on hook is active (see scope below).

Scope: what you touch and what you leave alone

Rewrite prose. Do not rewrite code.

• Rewrite: sentences and paragraphs meant for a reader. Documentation, articles, emails,

chat messages, commit descriptions written as prose, marketing copy, comments that are full sentences.

• Leave exactly as is: source code, code blocks, command lines, file paths, identifiers,

API names, configuration, log output, tool output, commit subject lines, and anything where a changed character breaks meaning or syntax. Technical precision wins. A function named delveIntoTree keeps its name.

When a passage mixes prose and code, rewrite the prose around the code and leave the code untouched.

Voice calibration (optional but better)

If the author gives you a sample of their own writing, or you have earlier text from them, read it first. Note:

• Typical sentence length and how much it varies.
• Words they reach for and words they avoid.
• Punctuation habits.
• Recurring turns of phrase.

Then rewrite toward that voice, not toward a generic "good writing" default. If you have no sample, aim for plain, direct prose and say nothing flowery.

The two-pass procedure (this is the QA step, and it is mandatory)

Never return a first draft. Always run both passes.

Pass 1, draft rewrite. Read the text. Find every pattern from the checklist below. Rewrite to remove them while keeping the meaning. Produce a draft.

Pass 2, self-audit. Read your own draft as a skeptic. Ask one question: "What in this still reads as AI?" Walk the full checklist again against your draft, not against the original. Fix what you find. Then do the dash sweep: search the draft for every em dash (Unicode U+2014) and en dash (Unicode U+2013) and remove all of them, replacing each with a period, comma, colon, parentheses, or a reworded sentence. The final output contains zero em dashes and zero en dashes.

Return the final version, then a short audit summary: which patterns you found and what you did. Keep the summary to a few lines. Do not pad it.

The checklist

Walk all four groups. Full examples for each ID are in [docs/patterns.md](../../docs/patterns.md).

Content

• C1 Inflated significance ("stands as a testament", "pivotal moment").
• C2 Notability name-dropping ("featured in numerous outlets").
• C3 Superficial -ing analysis (trailing "highlighting", "underscoring").
• C4 Promotional language ("nestled", "vibrant", "boasts").
• C5 Vague attribution ("experts argue", "studies show" with no source).
• C6 Formulaic challenge-and-future section ("despite challenges, it thrives").
• C7 Treating a common phrase as a proper name.

Language

• L1 AI vocabulary clusters (delve, tapestry, testament, underscore, intricate, leverage,

realm, navigate, robust, seamless, multifaceted).

• L2 Copula avoidance ("serves as", "boasts", "represents" for plain "is" or "has").
• L3 Negative parallelism ("not only X but Y", "it's not X, it's Y", "X rather than Y").
• L4 Rule of three (forced triples).
• L5 Elegant variation (synonym cycling for one thing).
• L6 False range ("from X to Y" with no real scale).
• L7 Hidden actor (passive voice, subjectless fragments).

Style and formatting

• S1 Title Case headings (use sentence case).
• S2 Boldface overuse.
• S3 Inline-header lists ("Label: sentence" repeated).
• S4 Em dash and en dash overuse. Hard cut, remove every one.
• S5 Curly quotes (use straight quotes).
• S6 Emoji as formatting.
• S7 Skipped heading levels.
• S8 Thematic break before a heading.

Tone and communication

• T1 Chat-assistant artifacts ("I hope this helps", "Let me know if").
• T2 Knowledge-cutoff disclaimers.
• T3 Sycophancy ("Great question!").
• T4 Filler phrases ("in order to", "due to the fact that").
• T5 Stacked hedging ("could potentially possibly").
• T6 Didactic disclaimers ("it's important to note").
• T7 Formulaic upbeat conclusion ("the future looks bright").
• T8 Signposting ("let's dive in").
• T9 Persuasive authority tropes ("at its core").
• T10 Theatrical openers ("Honestly?", "Here's the thing").

For pasted or imported text, also check the markup tells M1 to M4 in the catalog (stray model citation tokens, markdown in the wrong format, tracking parameters, broken citations).

Do not over-correct

Density is the signal, not any single word. Leave these alone:

• Clean grammar, formal register when the topic earns it.
• One uncommon word, or one dash the author clearly chose, with no other tells nearby.
• A key term repeated on purpose for clarity.
• Real voice: fragments, asides, a rhetorical question.

If a rewrite would strip the personality out of writing that already sounds human, stop. The goal is text that sounds like the author, not text scrubbed flat.

Output format

1. The final rewritten text.
2. A short audit summary: the patterns you found and removed, by ID or name. A few lines, no more.

If you changed nothing because the text already reads human, say so and explain why in one line.

Worked example

Input > Nestled at the intersection of design and engineering, our platform serves as a testament to > what's possible when teams delve into the intricate tapestry of modern workflows. It's not just > a tool, it's a complete reimagining. Experts agree the possibilities are endless.

Pass 1 draft > Our platform sits between design and engineering. Teams use it to manage modern workflows. It is > more than a single tool. Many people see real potential in it.

Pass 2 audit

• Still soft: "Many people see real potential" repeats the vague-attribution problem (C5) and the

upbeat-ending problem (C7, T7). Cut it or make it concrete.

• No dashes present. Good.

Final > Our platform sits between design and engineering. Teams use it to plan, assign, and track work in > one place. In our first quarter, 40 teams moved their workflows onto it.

Audit summary: Removed C4 (nestled), C1/L2 (serves as a testament), L1 (delve, intricate, tapestry), L3 (not just X, it's Y), C5 (experts agree), T7 (possibilities are endless). Added a concrete number in place of the vague close.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: badrusiddique
• Source: badrusiddique/ai-naturalize-skill
• License: MIT

Install and usage instructions live in the source repository linked above.