Less AI Tone Uses Data to Flag the Sentences That Sound Like AI

When a user has finished a Chinese draft (or had AI write one) and wants to remove the "obviously AI-written" feel, use this Skill: first run the script and semantic review to produce a detection report that lists the flagged sentences and suggested fixes, and rewrite only after the user replies, then re-test after editing and give a before-and-after comparison. It is built on the open-source lieflat-less-ai-tone (CC BY-NC 4.0: free for non-commercial use, contact the author for commercial use). Its rules come from statistical comparisons of thousands of human and AI articles, and it changes only phrasings measured to be clearly more common in AI text. Expressions that humans use just as often are never touched.

Triggers

The user says things like "remove the AI flavor", "this reads like AI wrote it", "help me make it sound human-written", or "check for AI traces", and provides a finished Chinese draft.

Hard rules

1. Report first, then rewrite. The first round delivers only the detection report. Do not touch the body text until the user replies "apply the suggestions" or specifies the scope of changes.
2. Minimal changes. For each flagged spot, change only what is necessary. Do not polish unprompted or swap out words that are fine.
3. Preserve information. Facts, numbers, names, quotations and qualifiers ("possibly" must not become "definitely") are all kept; every content word in the revised text must be traceable to the original.
4. Keep structure. Heading levels, paragraph order, lists, tables, quotations and code blocks stay as they are, and subheading numbering is not changed.
5. Do not create unevenness. Do not make sentences alternately very long and very short, or split paragraphs into fragments, in order to "sound human".
6. Read the style reference first if one exists. If the user's style document conflicts with this Skill, the style document wins (for example, if the author habitually uses dashes).
7. Never assert from detection results that the author used AI. The report only says "these phrasings are measurably more common in AI text".

Installation

npx skills add larashero3-dotcom/lieflat-less-ai-tone

Requires Python 3. Restart the Agent after installing. The commands below run inside the Skill directory; the detailed criteria are authoritative in the repository's SKILL.md, DETECT.md and references/semantic-criteria.md.

Workflow

1. Script detection (save the JSON before editing, since step 4 needs it):

   python3 scripts/detect.py draft.md --json > /tmp/before.json
   

2. Semantic review: extract candidates the script cannot count (synonym rotation, every paragraph tied to the previous one, conclusion-style colons, three sentences with identical structure), judge each one against the criteria in references/semantic-criteria.md, and write the results back to JSON:

   python3 scripts/detect.py draft.md --candidates --out /tmp/candidates.json
   

3. Deliver the detection report, then stop. The report should include: which indicators this article hit, how many instances of each, and how many are proposed for change versus kept; for each category, 1 to 2 quoted sentences from the original, an explanation of why it is worth changing, and how to change it; kept items with their reasons. Do not paste the raw script output to the user.

4. After the user replies, rewrite and produce a comparison report:

   python3 scripts/detect.py revised.md --json > /tmp/after.json
   python3 scripts/report.py draft.md revised.md /tmp/before.json /tmp/after.json --semantic /tmp/candidates_labeled.json --out ai-tone-report.md
   

   The default output is Markdown, with changes marked by strikethrough and insertion markers. For a web version, add --format html.

Phrasings to change (by priority)

Phrasing	Example	Fix
Dramatic reversal tone	"壁垒不是技术，而是认知" ("the barrier is not technology but cognition")	State the judgment directly: "壁垒是认知" ("the barrier is cognition")
Prompt-style colon	"一句话总结：成本太高" ("one-sentence summary: the cost is too high")	Delete the prompt label and write the statement directly; when the first half continues the previous text, only change the punctuation
Stock openers and translationese	Long preposed modifiers, Europeanized sentence structures	Split them up or convert to Chinese word order
Dense enumeration commas	Five or six parallel items in one sentence	Summarize if possible, or change the syntax of one item
Adjacent sentences with the same skeleton	Three sentences in a row of the form "A did X, B did Y, C did Z"	Break up one of them: merge, reorder, or shorten
Synonym rotation	"提升 / 优化 / 增强" ("raise / optimize / enhance") alternating within one paragraph	Standardize back to the original word
Every paragraph tied back	Each paragraph opens with "这也解释了为什么" ("this also explains why") to link back	Choose one or two places and switch topics directly
Dashes (varies by model)	Claude-style text uses many dashes	Handle by density; keep them if the author habitually uses them

Not changed

• Questions, rhetorical questions, and question-form subheadings (humans use these more than AI)
• Ordinal subheadings, "这意味着" ("this means") style restatements, and summarizing sentences (measured to be used equally or more by humans)
• Conversational, colloquial, and literary parallelism and metaphor
• Quotations, code, and foreign-language passages
• Whether a metaphor is apt: the model cannot judge this reliably, so it is only listed for the user to review and is not changed automatically

Output format

Round one: the detection report (indicator table, example sentences, fixes, and reasons for keeping items), ending with one sentence: "Reply 'apply the suggestions' or tell me which scope to adjust."

Round two: the full revised text plus the path to the before-and-after comparison report, and a note on re-test results: whether the flagged hits that needed handling have been cleared to zero, and which were kept and why.

In one sentence

Use data to point out which sentences sound like AI, wait for the user's approval, then change only those spots and leave everything else untouched, not a single word.