Levyra humanizer

Goal

Make finished prose sound like a person wrote it while preserving every factual claim. This is a lightweight final pass, not a research, review, or reasoning stage. Do not reopen the codebase or load extra evidence merely to humanize text.

Non-negotiable claim parity

• Keep every fact, number, date, path, version, test result, checkbox, citation, limitation, and risk state unchanged unless the source itself requires correction.
• Never turn UNRUN, BLOCKED, FAIL, uncertainty, or missing evidence into a stronger claim.
• Preserve required templates, headings, Markdown structure, code blocks, YAML, links, and release-note schema.
• Do not invent context, praise, motives, benefits, sources, or validation.

Natural prose pass

Prefer simple verbs, concrete nouns, varied sentence length, and the writer's existing vocabulary. When a real writing sample is available, match its tone and punctuation instead of imposing a generic style.

Remove only clear AI-writing residue:

• generic preambles, chatbot closers, fake-candid hooks, and "let me know" offers;
• inflated importance, sales language, vague expert claims, and generic optimism;
• filler, stacked qualifiers, forced groups of three, synonym cycling, and repeated headings;
• excessive bold mini-labels, decorative emoji, and formulaic "not X but Y" phrasing;
• em/en dashes when the writer's sample does not normally use them.

Keep technical language technical. Do not simplify a term when precision would be lost. One pattern alone is not a reason to rewrite natural prose.

Embedded PR/release mode

For PR descriptions, release notes, commit-adjacent prose, and other embedded uses:

1. work only from the already-drafted text and its verified evidence;
2. make one concise rewrite pass;
3. preserve all required Levyra template sections and unperformed checkboxes;
4. return only the final text unless the caller explicitly needs a change summary.

Do not spend engineering context on examples of writing style. The final prose pass should be much cheaper than the engineering work it describes.

Final check

Before returning, verify: same claims, same validation truth, same required structure, no invented facts, no lost limitation, and no obvious chatbot filler.

Provenance

This compact Levyra adaptation keeps the useful principles of the MIT-licensed blader/humanizer method while removing its large example catalog from the runtime skill body to reduce context cost.