Essay Humanize Iterator — Skill Specification
Purpose
Iteratively refine essays to minimize false positives from oversensitive AI detectors by removing stereotypical AI writing patterns and aligning semantic density and syntactic complexity with native human writing norms.
When to Use
- User submits an essay and wants to reduce AI stylistic patterns that trigger false positives
- User asks to rehumanize, iterate humanize, or improve writing naturalness
- User wants to improve semantic density or syntactic complexity to match human writing norms
- User mentions AI风格优化, 减少AI痕迹, 迭代改写, 写作自然度
Workflow
1. User provides essay text
2. MEASURE: Run skill/scripts/measure.py → get AI score, MDD, TTR, CW ratio
3. CHECK: If all metrics pass → output essay + report. Done.
4. REWRITE: Generate targeted revision using feedback from measurement
5. RE-MEASURE: Run measure.py on rewritten text
6. REPEAT: Loop steps 3-5 until pass or max iterations (default 3)
7. OUTPUT: Final essay + iteration report table + change summaryMeasurement Axes
| Axis | Tool | Pass Criteria |
|---|---|---|
| AI Pattern Score | 24-regex weighted scan | ≤ 15 / 100 |
| MDD Mean | spaCy dependency parse | 2.15 – 2.55 |
| MDD Variance | per-sentence MDD spread | ≥ 0.016 |
| Lexical TTR | content-word type/token | ≥ 0.50 |
| Content-Word Ratio | content / all tokens | 0.52 – 0.65 |
See skill/references/metrics.md for formulas and baselines.
Iteration Strategy
- Iter 1: Remove highest-weight AI patterns (em dashes, markdown, bolding, cliche metaphors)
- Iter 2: Fix remaining patterns + increase syntactic variety
- Iter 3: Fine-tune semantic density + register naturalness
See skill/references/iteration_strategy.md for full escalation logic.
Rewrite Engine
All rewriting is performed locally by the orchestrating LLM based on targeted feedback from measure.py. No external API calls are made.
Rules for rewriting:
- Process the essay paragraph by paragraph
- Follow the specific feedback instructions from
build_iteration_feedback() - Preserve all citations, references, and factual claims
- Do not add new sources or fabricate evidence
- Output plain text only (no markdown formatting, no LaTeX delimiters)
Output Format
Final Essay
Plain text. Preserve the original heading structure if any. No markdown artifacts.
Iteration Report
| Iter | AI Score | MDD Mean | MDD Var | TTR | CW Ratio | Status |
|------|----------|----------|----------|--------|----------|--------|
| 0 | 45.2 | 2.4821 | 0.0098 | 0.4712 | 0.6280 | FAIL |
| 1 | 18.6 | 2.3891 | 0.0142 | 0.4988 | 0.5932 | FAIL |
| 2 | 11.3 | 2.3504 | 0.0178 | 0.5124 | 0.5801 | PASS |Change Summary
After the table, provide a brief bullet list of what changed across iterations:
- Which patterns were removed
- How sentence structure was varied
- What vocabulary changes were made
Rules
- Preserve argument: The author's thesis, evidence, and logical flow must remain intact
- Preserve citations: Never remove, alter, or fabricate citations/references
- Plain text output: No markdown headings (unless input had them), no bold, no em dashes
- No hallucination: Do not add claims, data, or sources not in the original
- Idempotent measurement: Always use
measure.pyfor scoring — do not estimate scores - Early exit: If the input essay already passes all thresholds, output it unchanged with a passing report
- Transparency: Always show the iteration table so the user sees the convergence trajectory
Supporting Files
| File | Purpose |
|---|---|
skill/scripts/measure.py | Quantitative scorer (AI patterns + MDD + semantic density) |
skill/scripts/iterate.py | Iteration engine (measure + feedback generation) |
skill/references/patterns.md | 24 AI pattern definitions and fix strategies |
skill/references/metrics.md | Metric formulas, baselines, thresholds |
skill/references/iteration_strategy.md | Per-iteration focus and escalation logic |
data/analysis/weights.json | Corpus-derived pattern weights |