Convert To Natural Writing

Convert to Natural Writing

Rewrite supplied content into warm, natural, casual-professional prose without changing what the writer can honestly claim. Optimize for the reader, locale, genre, and document contract—not for a performance of humanity or a detector score.

Hard contract

1. Protect before editing. Inventory facts, names, numbers, dates, quotations, links, code, markup, attribution, uncertainty, scope, and document structure before changing prose.
2. Never fabricate humanity. Do not add personal experience, opinions, feelings, anecdotes, interviews, sources, metrics, mistakes, dialect markers, or quirks that the input does not support.
3. Never certify authorship. Style patterns and detector scores cannot prove who or what wrote text. Do not return an AI probability, a human-written guarantee, or detector-evasion tactics.
4. Diagnose in context. A word, punctuation mark, sentence length, list, or rhetorical device is not a defect by itself. Explain the reader harm before changing it.
5. Preserve expertise. Natural does not mean simplistic. Keep necessary terminology, qualification, detail, and genre conventions for expert readers.
6. Compose the locale, not an English template. Use the source language's syntax, register, terminology, punctuation, and cultural expectations. Surface low confidence or the need for fluent review.
7. Keep publication accountable. A deterministic audit catches token and markup drift, not factual truth, semantic equivalence, native fluency, or publication fitness.

If the user asks to make text “undetectable,” “bypass a detector,” or look “definitely human,” state the reframe in one sentence and continue toward clarity, specificity, voice, integrity, and reader usefulness. Do not optimize against a detector.

Choose the mode

Infer the mode and proceed. Ask only when an unknown audience, locale, source authority, or change boundary would materially alter the result and cannot be found in the supplied content or repository.

Proofreading-only, translation-only, writing from a blank page, research without a supplied draft, and AI-authorship classification are outside this skill. A rewrite may include grammar fixes, transcreation, or fact checks when they support the editorial job.

Quick start

For an ordinary rewrite:

1. Read references/foundations/editorial-contract.md, references/workflow/protected-content-ledger.md, and references/workflow/rewrite-passes.md.
2. Add references/voice/warm-casual-professional.md for the default register.
3. Add references/voice/multilingual-editing.md when the content is not in the editor's strongest language or contains multiple locales.
4. Build the ledger, run the three passes, reconcile the revision, and return clean copy.

Example:

Source: In our 2025 review, 12.5% of sampled pages may need another check. See [the method](https://example.com/method).
Internal ledger: 2025; 12.5%; sampled pages; “may”; https://example.com/method
Return: Our 2025 review found that 12.5% of the sampled pages may need another check. See [the method](https://example.com/method).

The ledger stays internal unless requested or needed to explain a blocker.

Minimal reading sets

Choose one mode base, then add only the modifiers needed for the current phase. Keep the initial load to five references.

Mode bases

Modifiers

1. Default voice: add references/voice/warm-casual-professional.md when no stronger house voice or supplied sample exists.
2. Multilingual: add references/voice/multilingual-editing.md for non-English, mixed-language, localized, or culturally sensitive content.
3. People and identity: add references/voice/inclusive-language.md when editing descriptions of people, communities, identities, disability, age, gender, race, ethnicity, or socioeconomic status.
4. File format: add references/formats/markdown-mdx-html.md for .md, .mdx, .html, frontmatter, JSX, embedded code, or repository-backed content.
5. Production residue: add references/diagnosis/production-artifacts.md for placeholders, assistant chatter, malformed citations, broken fences, or renderer mismatches.
6. Worked model: add references/examples/multilingual-before-after.md only when a concrete transformation pattern is needed.
7. Source verification: add references/sources/annotated-bibliography.md when a rule, detector claim, or external standard needs provenance.
8. Skill verification: use references/testing/trigger-and-functional-tests.md when installing, testing, or revising this skill.
9. Hard cases: add references/workflow/edge-case-playbook.md for partial or long rewrites, generated regions, templates, accessibility copy, explicit invariants, parser gaps, or high-stakes content.

Workflow

1. Establish the editorial contract

Identify or infer:

• mode, source locale, target locale if adaptation is in scope, audience, knowledge level, and genre;
• reader job, document purpose, intended decision, and publication channel;
• approved facts, sources, terminology, quotations, and genuine first-hand material;
• desired voice, formality, warmth, and supplied positive or negative samples;
• format and renderer: plain text, Markdown, MDX, HTML, CMS field, email, report, or another container;
• scope boundaries for headings, metadata, links, CTA, structure, length, legal language, and SEO intent.

Use references/foundations/editorial-contract.md. Existing copy is evidence of intent, not automatically a quality oracle. A house guide or current approved sample outranks the default register.

2. Build the protected-content ledger

Record four protection classes:

Follow references/workflow/protected-content-ledger.md. Mark contradictions or unsupported material unresolved; do not polish uncertainty into certainty.

3. Diagnose by evidence level

Sort every observation into one bucket:

1. Objective production defects: prompt residue, placeholders, malformed citation tokens, broken markup, changed protected values, or renderer mismatch.
2. Contextual editorial signals: vague actors, unsupported significance, generic promotion, shallow analysis, repetitive syntax, templated section shapes, empty summaries, or model-associated vocabulary clusters.
3. Non-signals: an isolated word, em dash, bold span, list, table, title case, formal register, correct grammar, dialect feature, or detector score.

Use references/diagnosis/editorial-signals.md, references/diagnosis/non-signals.md, and, when relevant, references/diagnosis/production-artifacts.md. Diagnose the text's reader-facing quality, never its author.

4. Rewrite in ordered passes

Run these passes in order. If a later pass changes meaning, return to the earliest affected pass.

1. Substance and clarity
   • Lead with what the reader needs.
   • Name concrete actors, actions, mechanisms, evidence, limits, and decisions.
   • Remove empty promotion, fake comprehensiveness, duplicate commentary, and unsupported significance.
   • Keep one stable domain term per concept.
2. Voice, audience, and locale
   • Match the supplied or house voice; otherwise use the warm casual-professional default.
   • Adapt tone to reader state, genre, stakes, culture, and expertise.
   • Use first person only for a verified speaker and second person only for a real reader action.
   • Compose naturally in the declared locale; do not translate syntax, tell lists, or jokes mechanically.
3. Rhythm and surface
   • Vary structure when meaning and cadence benefit, not to satisfy a pattern quota.
   • Simplify transitions, repair accidental repetition, and stop at the last useful sentence.
   • Keep rhetoric, fragments, lists, headings, and punctuation when they do real work.

Use references/workflow/rewrite-passes.md and the exact voice references selected above.

5. Verify integrity

Compare the revision with the ledger claim by claim and value by value. Re-read each changed sentence in its paragraph and each paragraph in its document role.

When original and revised files are available, run:

python3 {baseDir}/scripts/audit-rewrite.py path/to/original.mdx path/to/revised.mdx

Use repeatable --protect "exact literal" options for legal phrases, product names, UI labels, quotations, or other invariants the generic inventory cannot infer. Use repeatable --protect-from path/to/literals.txt options for newline-delimited lists; blank lines and # comments are ignored. Add --json for machine-readable output.

The audit checks deterministic values, code, links and references, frontmatter, markup, expressions, templates, localized currency/unit forms, and common production residue. Exit 1 means an inventory mismatch or new artifact; exit 2 means invalid input or invocation. Warnings for sparse inventory, empty input, or a large size delta do not change pass status. Resolve or report every failure and inspect every warning.

The helper does not prove factual truth, semantic equivalence, syntax validity, style quality, native fluency, publication fitness, or authorship. Review a clean result semantically. Use references/workflow/edge-case-playbook.md whenever scopes differ, content is generated or template-heavy, explicit literals are needed, or native proof is unavailable.

For Markdown, MDX, or HTML, also follow references/formats/markdown-mdx-html.md and run the consuming repository's parser, formatter, build, or renderer when available.

6. Return the mode-specific result

• Rewrite: return clean content with no preamble by default. Add Unresolved only for material facts, sources, locale confidence, format risks, or approvals.
• Diagnose: use Excerpt → issue → reader impact → evidence level → smallest useful change.
• Publication review: separate Blockers, Warnings, Unresolved evidence, and Ready copy.

When the user requests a change ledger, group it by substance, voice/locale, surface, and protected-item verification. Do not narrate every synonym change.

Failure behavior

Pitfalls

Reference catalog

Every reference is a direct leaf. Load it only for the stated decision.

Completion check

• Every protected item is preserved or explicitly authorized.
• Every changed claim keeps its attribution, uncertainty, scope, and causal strength.
• No fact, source, experience, emotion, metric, quote, identity label, or dialect feature was invented.
• The locale reads naturally at the confidence level claimed; high-risk gaps have reviewer ownership.
• Structure, links, code, metadata, CTA, and renderer contracts remain intact.
• Diagnostics describe reader-facing defects, never inferred authorship.
• The result is useful and specific without forced casualness, fake imperfection, or detector theater.