scholar-anti-ai-writing

Scholar Anti-AI Writing

Detect and eliminate AI writing patterns from academic prose. Based on Wikipedia's "Signs of AI writing" guide. Supports both English and Chinese text.

Core insight: LLMs predict the most statistically likely outcome — creating detectable patterns. This skill identifies and rewrites those patterns while preserving meaning and adding authentic voice.

Pre-load (before execution)

1. Codebase docs: If .workflow/codebase/ARCHITECTURE.md exists, read for project context
2. Specs: maestro load --type spec --category coding — load coding conventions
3. Wiki knowledge: maestro search "academic writing research paper" --json — top 5 entries as prior context
4. All optional — proceed without if unavailable

Architecture Overview

User: "Remove AI patterns from my paper"
         |
         v
┌──────────────────────────────────────────────────────────────┐
│  SKILL.md (Orchestrator)                                      │
│  Collect preferences → Dispatch phases → Track progress       │
└──────────┬───────────────────────────────────────────────────┘
           |
   ┌───────┼───────────┬──────────────┐
   v       v           v              v
┌──────┐┌──────────┐┌──────────────┐
│ P1   ││ P2       ││ P3           │
│Detect││ Rewrite  ││ Validate     │
│ &    ││ & Polish ││ & Score      │
│Score ││          ││              │
└──┬───┘└────┬─────┘└──────┬───────┘
   │         │             │
   v         v             v
 pattern   polished     final score
 report    prose        + diff report

Key Design Principles

1. Pattern detection + soul injection: Removing AI patterns is only half the job — add authentic voice.
2. Preserve meaning: Core content and technical claims must remain intact.
3. Language-aware: English and Chinese have different AI pattern signatures.
4. Scoring-driven: Quantitative 5-dimension scoring (50-point scale) guides revision.
5. Academic tone: Maintain scholarly register — humanize without becoming informal.

Interactive Preference Collection

Before dispatching to any phase, collect these preferences:

Questions to ask the user:

1. Input Source
   "What text should I process? (file path, directory, or paste text)"
   → inputSource

2. Language
   Options: English | Chinese | Bilingual (auto-detect per section)
   → language

3. Writing Context
   Options: Academic Paper | Technical Report | Blog/Article | General Prose
   → writingContext

4. Aggressiveness
   Options: Conservative (minimal changes) | Balanced (Recommended) | Aggressive (heavy rewrite)
   → aggressiveness

Store responses as antiAiPreferences context for all phases.

Auto Mode Defaults

When workflowPreferences.autoYes === true:

• Process all .tex files in cwd
• Auto-detect language
• Academic Paper context, Balanced aggressiveness
• Skip confirmation prompts

Execution Flow

COMPACT DIRECTIVE: Context compression MUST check TodoWrite phase status. The phase currently marked in_progress is the active execution phase — preserve its FULL content. Only compress phases marked completed or pending.

TodoWrite Setup

Anti-AI Writing:
- [ ] Phase 1: Detect & Score — scan for AI patterns, generate initial score
- [ ] Phase 2: Rewrite & Polish — rewrite flagged sections, add voice
- [ ] Phase 3: Validate & Score — re-score, verify quality threshold

Phase Sequence

Phase 1: Detect & Score
   └─ Ref: phases/01-detect-score.md
      ├─ Input: inputSource, language
      └─ Output: patternReport (flagged passages + initial scores)

Phase 2: Rewrite & Polish
   └─ Ref: phases/02-rewrite-polish.md
      ├─ Input: patternReport, writingContext, aggressiveness
      └─ Output: polishedText (rewritten content)

Phase 3: Validate & Score
   └─ Ref: phases/03-validate-score.md
      ├─ Input: polishedText, original text
      └─ Output: finalReport (before/after scores, diff)

Phase Reference Documents (read on-demand when phase executes):

Compact Rules:

1. TodoWrite in_progress → preserve full content, do not compress
2. TodoWrite completed → may compress to summary
3. sentinel fallback → phases marked with sentinel contain compact sentinel; if only sentinel remains, must immediately Read() to recover

Core Rules

1. Never lose meaning: Technical claims, data, and specific facts must survive rewriting.
2. Score before and after: Every processed text gets a 5-dimension score (50-point scale).
3. Flag, don't force: In Conservative mode, highlight issues but let the user decide.
4. Academic register: For papers, maintain formal-but-natural tone. No slang or casual language.
5. Minimum threshold: Target score >= 35/50 for academic papers, >= 40/50 for submission-ready.

Quick Scoring System (5 dimensions, 10 points each)

Data Flow

Phase 1 ──patternReport──→ Phase 2
Phase 2 ──polishedText──→ Phase 3
Phase 1 ──originalText──→ Phase 3 (for comparison)

Error Handling

Related Skills

• scholar-writing: Phase 5 (Anti-AI Polish) uses this skill's patterns
• scholar-review: Uses scoring to assess writing quality