domains / media / de-ai-rewrite / de-ai-writer

de-ai-writer

Skill card

combined

The two copies overlap on the goal (removing AI smell from Chinese copy) but each contains material steps the other lacks. The davepoon copy is a thin wrapper around a local Python script (deai.py) with API-key and prompt-only modes. The sickn33 copy is a documentation-only procedure that defines a deterministic AI-smell scoring formula, a 35-pattern catalog, a diagnose-rewrite-verify loop, and optional style-clone/variant/tone steps. Crowning one would hide a real alternative: a script-based tool versus a manual, formula-driven editing procedure.

Repositories

Other copies

Combined skill

# De-AI Writer — Chinese AI-Smell Removal

Rewrite AI-generated Chinese copy into natural, human-sounding text. Use when the user asks 去AI味, 改得像人写的, or provides a draft that reads machine-written.

## Step 1: Diagnose — Score the Draft

1. Scan the text against the 35-pattern Chinese AI-tell catalog (e.g., 赋能/闭环, 首先-其次-最后, 随着…的发展, 拔高意义 endings, 公文套话).
2. Count hits per paragraph (repeats in the same paragraph count once).
3. Weight: strong patterns = 2 points; weak-evidence patterns (em-dash, hedges, passive/no-subject, 的-stacking, inconsistent quotation marks) = 1 point.
4. Normalize: D = weighted total / max(1, total characters / 100).
5. Index: AI味指数 = min(100, round(D × 10)).
6. Bands: 0–20 基本像人写 · 21–45 轻度 AI 味 · 46–75 明显 AI 味 · 76–100 一眼假.
7. Report three numbers: 命中处数 / 加权总分 / AI味指数.

## Step 2: Rewrite Against Specific Patterns

- Work through the hit list one pattern at a time.
- A single weak-evidence hit is not evidence; act only when two or more appear in the same paragraph.
- Delete, don't decorate: drop the negation half of 不是 X，而是 Y, drop significance endings, drop 开场铺垫, drop assistant residue.
- Facts come from the source only. Never introduce a fact, number, or specification the source does not contain. If a fact is needed for clarity, ask the user.

## Step 3: Verify the Rewrite

- Re-score the rewritten text with the same formula.
- Report the before/after triple (命中处数 / 加权总分 / 指数) and how many patterns were cleared.
- Confirm no fact was added or lost. A rewrite that lowers the score by deleting facts is a failed edit.

## Step 4: Optional Extensions

- **Style clone**: match a reference sample's sentence rhythm, vocabulary, and colloquial ratio.
- **Variants**: produce 2–6 clearly different versions for A/B testing.
- **Tone**: re-target to casual, formal, marketing, humor, or direct register.
- **Review**: score on Hook / Pacing / Emotion / AI-Smell / Clarity / Persuasion / Structure / Readability, plus three concrete improvements.

## Alternative: Script-Based Mode

If a local rule engine is available (e.g., scripts/deai.py from the source repository):

1. With an API key:
   ```bash
   export LLM_API_KEY="your-key"
   export LLM_BASE_URL="https://api.deepseek.com/v1"
   export LLM_MODEL="deepseek-chat"
   python3 scripts/deai.py draft.txt -o output.txt
   ```
2. Without an API key (zero-cost):
   ```bash
   python3 scripts/deai.py draft.txt --prompt-only
   ```
   Paste the output prompt into any AI assistant to perform the rewrite.

## Conflicts

- davepoon/buildwithclaude ships a runnable Python script as the primary workflow; sickn33/agentic-awesome-skills is documentation-only and does not include the script.
- sickn33/agentic-awesome-skills requires a deterministic scoring formula before and after rewriting; davepoon/buildwithclaude has no scoring step.
- davepoon/buildwithclaude offers a --prompt-only mode for external AI assistants; sickn33/agentic-awesome-skills does not mention this mode.