Strip the ai voice

strip-the-ai-voice

AI Chinese reads wrong because of what it adds. Don't rewrite it — delete it.

Three load-bearing positions

1. The AI voice is added, so the fix is deletion, not revision

• Rule: Never paraphrase a sentence to make it "sound natural". Remove the added material; keep what remains.
• Mechanism: Rewrite-style humanizing is unrepeatable — each pass can introduce new decoration, and the same input gives different outputs. Deletion is single-direction and checkable sentence by sentence: either a word was removed or it wasn't.
• Field observation: Archival documents written under a strict restraint standard, when run through "polish the wording" tools, regress — honorifics and adjectives come back. Only delete-only rules held the restraint across repeated edits.
• Boundary: If the text is missing information (not over-decorated), deletion cannot fix it. Report the gap; do not write new content to fill it.

2. The criterion is the delete-test, not taste

• Rule: Delete every adjective and adverb in the sentence. If the remainder stands, use the remainder. If the remainder is a stump, the whole sentence was decoration — delete the sentence.
• Mechanism: This turns the subjective question "is this restrained enough" into a binary, repeatable check. Two runs on the same text converge to the same output.
• Field observation: "读者将会深刻地感受到机制的精妙之处" → after deletion: "读者会感受到机制" → a stump → delete the whole sentence. The sentence carried no information; the test exposed it.
• Boundary: The test judges decoration, not correctness. A restrained sentence can still be wrong; that is outside this skill.

3. Touch only AI-generated copy. Never the user's prose

• Rule: This skill operates on copy the tool itself produced — labels, prompts, captions, sign-offs, boilerplate. The user's own writing is not edited, not even one character.
• Mechanism: Restraint is a discipline for what the tool says, not a license to override the author. Editing the author's prose is a different act with different stakes — and a deletion there is irreversible.
• Field observation: This line recurs in the source protocols this skill was distilled from: "只管你自己生成的文案,正文一个字不动". The skills that held that line stayed usable; mixing the two scopes is where trust breaks.
• Boundary: If the user explicitly hands over their own prose and asks for this treatment, the scope changes by their instruction — state the new scope back before deleting.

Map

Task	Load
What to delete (four classes, word lists)	reference/delete-list.md
The delete-test; delete-word vs delete-sentence line	reference/the-test.md
Hard forbidden items; symbol rule (structural stays, emotional goes)	reference/forbidden.md
✗→✓ before/after table	reference/examples-table.md
Full worked example	examples/before-after.md
Adaptation lineage and test record (vs humanizer-zh)	examples/lineage-test.md

Workflow gate

Before deleting anything, state in one line: which spans are AI-generated copy to be processed, and where the user's own prose begins. If that boundary cannot be drawn, stop (see Failure behavior).

Fields

• Version: 0.1 (2026-06). Changelog: corrections are marked in place, not erased.
• Sunset condition: Retire when base models stop emitting honorific- and decoration-laden Chinese by default, and "去 AI 味" stops recurring as a request.
• Failure behavior: Cannot tell AI copy from user prose → stop and ask. Never delete text that might be the user's original (irreversible). No silent best-effort output.
• Composition: composes-with any Chinese writing skill. conflicts-with: humanizer-zh — opposite direction (it adds human texture; this deletes decoration). Do not apply both to the same text. Explicit user instruction overrides this skill.
• Time: Judgment fixed 2026-06. The default "AI flavor" shifts with model generations; review in 12–18 months.
• Boundary conditions: Validated on professional, formal, and archival Chinese copy. Not for marketing copy, emotional writing, or any text where spoken warmth is the point — there, deletion makes it worse, and this skill should stay silent.