deslop tells you "how to revise"; this skill tells you "how severe it is, which items are deducted, and which sentence the problem is in". Turn "I feel like there's a bit of AI flavor" into a quantifiable diagnosis:AI-flavor index 7.2/10, deductions on metaphor density and consecutive same sentence patterns, locations see list.
When to use
- Wants to give a passage/chapter of main text an "AI-flavor score"
- Before revising, first locate the worst AI-flavor areas and cut by priority
- Quantitatively compare how much "before vs after revision" was reduced
- Division of labor with adjacent skills:
aidetect：Quantitative detection + scoring + locating(produces diagnosis)deslop: surface revision (rhetoric/punctuation/repetition/extreme words)human: deep revision (loss of control/irrationality/non-closure)annotate: single-chapter rhythm visualization (hook/payoff/modifier words)review: five-dimension diagnosis of reader retention for a whole volume
Twelve quantitative indicators (see details in references/ai-detector.md）
- Metaphor density ≤3 per 1,000 characters and not used consecutively ("like… as if… just like…" three in a row = contamination)
- Parallelism/three-part pattern ≤1 place/chapter ("not A, not B, but C")
- Consecutive same sentence patterns ≤2 sentences (3+ consecutive sentences of "subject + action" = warning)
- Sentence-length fluctuation Within one paragraph, long/short ≥3 times; need both fragments and long sentences
- Paragraph distribution Long and short paragraphs alternate; avoid all-long/all-short/identical parallelism
- High-frequency emotion word repetition shocked/pupils contract/heart jolts… each ≤1 per chapter
- Extreme words very/extremely/instantly/suddenly… delete ninety percent, ≤2–3 per 1,000 characters
- Transition words/empty elevation/misused punctuation however/at the same time…、
.../period used as enumeration comma - Summary sentences / tell density "In this moment she understood…" "she was very sad" → show≫tell (for this specific item switch to
emotion） - Action/dialogue/psychology/exposition ratio Prevent "80% psychology"; psychology + exposition must not exceed half
- Environment audit Environment ÷ progression; 500 characters of blue sky and white clouds without advancing plot = padding
- Reading friction Consecutive long paragraphs/environment/recall/explanation stacked → drives mobile readers away
⚠️ This isan estimation-style heuristic rule, not word-by-word counting: scan once to estimate occurrences per 1,000 characters, find consecutive sections, and pick the densest locations. Focus onlocating + trends。
AI-flavor index (0–10, higher = stronger = more in need of revision)
Consecutive same sentence patterns and sentence-length fluctuation carry the highest weight (they most destroy the human feel). Tiers:
- 0–3 like a real person: occasional minor flaws, publishable
- 4–6 half-finished: has machine traces, revise key items according to locations
- 7–10 strong AI flavor: obviously fake at a glance, run the whole chapter through
deslop+human
Output: quantitative scan report
【AI 味量化扫描 · 第 N 章 / 选段】
字数：__ AI 味指数：__/10 一句话结论：……
指标(实测 vs 阈值)：比喻密度 / 连续同句式 / 句长波动 / 段落分布 /
高频情绪词 / 极端词 / 排比三段式 / 转折升华标点 —— 逐项 ✓/⚠/✗
【扣分定位】（给到句/段）
① 第_段："原句" → 连续4句主语+动作 → 改法
② 第_段："像…似…宛如…" → 比喻三连 → 删两个
【优先改】先动 AI 味分最高的 1–2 项
Principles
- Locate first, then score: Scores serve "what to change", not scoring for its own sake.
- Consider the platform: Tomato is inherently short-sentence and fast-paced—dense short sentences ≠ AI; the key iswhether there is variation, whether fragmented sentences and long sentences are interspersed。
- Diagnose only, don't revise: After the report is produced, hand it to
deslop/human/emotionto do the work.
Coordinate with
- Surface changes →
deslop（references/anti-ai-checklist.md） - Deep human feel →
human（references/human-texture.md） - Sentence rhythm →
references/sentence-rhythm.md - Emotions acted out →
emotion（references/show-emotion.md）