prose-polish（散文润色）

Prose Polish v2

Evaluate and elevate writing effectiveness through multi-dimensional quality assessment. Goal is not "less AI-like" but genuinely better writing—coherent, credible, purposeful, and distinctive.

Philosophy

Writing Effectiveness = f(Text, Author, Audience, Context, Genre)

We optimize for quality, not undetectability. These often correlate, but the distinction matters:

• Bad goal: "Make this not sound like AI"
• Good goal: "Make this effective writing"

Quick Start

Analysis: Detect genre → Load detection-patterns.md → Apply 6-dimension evaluation → Generate quality profile

Elevation: Analyze → Load remediation-strategies.md → Phase 1 (Structure) → Phase 2 (Style) → Explain changes

Prevention: Load prevention-prompts.md → Build genre-calibrated constraints → Generate → Self-verify

Core Capabilities

1. Detection & Analysis

When: User asks to "analyze," "evaluate," "check," or "score" text

Process:

1. Detect Genre (before scoring)

- Technical | Business | Academic | Creative | Personal | Journalistic - Apply genre-appropriate thresholds

1. Load references/detection-patterns.md

1. Perform 6-Dimension Analysis:

- Craft (0-100): Lexical patterns, structural variance, rhetorical execution - Coherence (0-100): Logical flow, functional specificity, earned transitions - Authority (0-100): Earned vs delegated vs false expertise signals - Purpose (0-100): Clear intent, stakes, audience calibration - Voice (0-100): Distinctiveness, embodiment, appropriate register - Effectiveness (0-100): Genre-weighted synthesis

1. Generate Quality Profile

Report Format:

WRITING EFFECTIVENESS ANALYSIS

Genre: [Detected] | Calibration: [Applied]

QUALITY PROFILE:
           Craft: ████████░░ 80 - [Brief interpretation]
       Coherence: ██████░░░░ 60 - [Brief interpretation]
       Authority: █████░░░░░ 50 - [Brief interpretation]
         Purpose: ███████░░░ 70 - [Brief interpretation]
           Voice: █████████░ 90 - [Brief interpretation]
   Effectiveness: ███████░░░ 70 - [Genre-weighted average]

KEY INSIGHT: [Diagnostic based on dimension gaps]
Example: "High craft but low authority = generic specificity problem"

DETAILED ANALYSIS:

CRAFT ISSUES:
- Lexical: [specific patterns, with genre context]
- Structural: [sentence variance, paragraph patterns]
- Rhetorical: [commitment level, specificity quality]

COHERENCE ISSUES:
- Logical flow: [do ideas connect across paragraphs?]
- Specificity function: [relevant vs decorative details]
- Transition authenticity: [earned vs mechanical]

AUTHORITY ISSUES:
- Type: [Earned / Delegated / False / Mixed]
- Expertise signals: [insider knowledge present/absent]
- Stakes: [skin in the game visible?]

PURPOSE ISSUES:
- Intent clarity: [what is this FOR?]
- Audience calibration: [appropriate for reader?]
- Stakes: [why should reader care?]

VOICE ISSUES:
- Distinctiveness: [recognizable author?]
- Embodiment: [feels like a person?]
- Register: [appropriate for genre?]

TOP 5 PRIORITY IMPROVEMENTS:
1. [Most impactful, actionable fix]
2. [...]
3. [...]
4. [...]
5. [...]

Scoring Philosophy:

• Be ruthless in scoring. Avoid grade inflation.
• Dimension gaps are diagnostic (high craft + low coherence = decorative writing)
• Genre calibration prevents false positives on appropriate conventions

2. Elevation & Remediation

When: User asks to "improve," "fix," "elevate," or "rewrite" text

Process:

1. Perform quick 6-dimension analysis
2. Load references/remediation-strategies.md
3. Apply Two-Phase Remediation:

Phase 1: Structural (The Editor) Focus on logic and authority before touching style.

• Coherence Pass:

- Check: Does logic flow across paragraphs? - Check: Is every detail doing work? - Fix: Remove decorative specificity - Fix: Repair logical gaps - Fix: Ensure transitions are earned

• Authority Pass:

- Check: Is authority earned or delegated? - Fix: Replace institutional voice with speaker - Fix: Add demonstrated expertise signals - Fix: Introduce appropriate stakes/vulnerability

Phase 2: Stylistic (The Writer) Now refine rhythm, commitment, and voice.

• Rhythm Pass:

- Sentence variance per genre threshold - Structural breaks appropriate to genre - Information density variance (avoid uniform medium-density)

• Commitment Pass:

- Remove cowardly hedges (opinion avoidance) - Preserve protective hedges (epistemic honesty) - Add functional specificity - Make claims with stakes

• Voice Pass:

- Add embodiment markers - Inject appropriate personality (avoid "LinkedIn Influencer" overcorrection) - Risk-taking calibrated to genre

Output:

ELEVATED VERSION:
[Rewritten text]

PHASE 1 CHANGES (Structure):
- Coherence: [What logical issues were fixed]
- Authority: [How expertise was demonstrated]

PHASE 2 CHANGES (Style):
- Rhythm: [Sentence variation details]
- Commitment: [Hedge removal, specificity additions]
- Voice: [Personality calibration]

BEFORE/AFTER EXAMPLES:
[3-5 transformations with principles explained]

Depth Control (Aggressiveness Levels):

Users can control the extent of remediation:

How to request:

• "Fix the logic but keep my voice" → Conservative
• "Improve this while keeping the general tone" → Moderate
• "Rewrite this for maximum effectiveness" → Aggressive

Default: Moderate (both phases, respects original intent)

3. Prevention & Generation

When: User asks to "write" or "generate" with quality emphasis

Process:

1. Identify genre and audience
2. Load references/prevention-prompts.md
3. Construct genre-calibrated constraints
4. Generate with quality dimensions in mind
5. Self-verify against 6-dimension framework
6. Refine if any dimension scores below threshold

4. Training & Teaching

When: User wants to learn quality evaluation

Process:

1. Load appropriate reference files
2. Explain the 6 dimensions and why they matter
3. Show examples of dimension gaps (high X, low Y)
4. Demonstrate genre calibration effects
5. Practice exercises with real text

Genre Calibration

Detect genre before scoring. Apply appropriate thresholds:

Genre-Specific Signals

Technical Documentation:

• Allow: "certain," "particular," "specific" (precision, not hedging)
• Allow: Consistent sentence length (clarity, not robotic)
• Require: Explains WHY not just HOW
• Authority: Demonstrated through insider terminology and tradeoff awareness

Business Writing:

• Require: Friction acknowledgment (what challenges exist?)
• Require: Clear ownership and next steps
• Watch: Institutional hiding ("it is recommended" vs "I recommend")
• Authority: Numbers with interpretation, not just data dumps

Academic Writing:

• Require: Synthesis over summarization
• Require: Clear contribution statement
• Allow: "It appears that" as epistemic honesty
• Authority: Citation genealogy, not just name-dropping

Creative/Narrative:

• Require: Surprise, sensory embodiment
• Require: Specificity that reveals character, not decorates
• Watch: Generic emotional beats ("hollow ache" without texture)
• Authority: Earned through embodied experience

Dimension Deep Dives

Coherence (NEW in v2)

What it catches: Decorative specificity, logic gaps, non-sequiturs

Red Flags:

• Details that don't advance understanding
• Causal claims that don't hold ("teaching calculus → cracked hands")
• Transitions that connect syntactically but not semantically
• Specificity that signals "human-ness" rather than builds meaning

Questions to Ask:

1. If I remove transitions, do ideas still connect?
2. Could I swap paragraphs without changing meaning? (Bad if yes)
3. Is every specific detail doing work?
4. Would a hostile reader find logical gaps?

Authority (NEW in v2)

What it catches: Performed expertise vs demonstrated expertise

Authority Types:

• Earned: Insider details, vulnerability, consequences for being wrong
• Delegated: Citations without synthesis, institutional voice, numbers without interpretation
• False: Stereotypes as expertise, generic specificity, authority cosplay

Note: We measure *signaling*, not *truth*. An LLM cannot verify facts—it can only assess whether authority markers are present. Be honest about this limitation.

Hedge Classification (NEW in v2)

Not all hedges are bad. Classify before penalizing:

Cowardly Hedges (PENALIZE):

• Avoiding opinion: "Some might say," "It could be argued"
• Diluting claims: "somewhat," "fairly," "rather"
• Escape hatches: "in a sense," "in many ways"

Protective Hedges (PRESERVE):

• Epistemic honesty: "The evidence suggests," "Current research indicates"
• Appropriate uncertainty: "appears to," "likely"
• Precision: "certain," "particular," "specific"

Quality Standards

Every output should pass:

Coherence Test: Do ideas connect logically across the piece? Authority Test: Is expertise demonstrated, not just claimed? Purpose Test: Is it clear what this writing is FOR? Voice Test: Would you recognize this author again? Embodiment Test: Does it feel like a person with experiences? Surprise Test: Is there at least one unexpected element?

Effectiveness Calculation

Genre-weighted average:

Reference Files

When to Load:

• detection-patterns.md: Always for analysis (includes coherence patterns, authority signals, hedge classification)
• remediation-strategies.md: Always for elevation (two-phase framework)
• prevention-prompts.md: Always for generation (genre-calibrated constraints)
• ai-vocabulary.md: When checking specific banned words

Critical Principles

1. Effectiveness Over Detection: Goal is quality writing, not fooling detectors.

1. Genre Calibration Is Non-Negotiable: Technical docs and creative fiction have different rules.

1. Dimension Gaps Are Diagnostic: High craft + low coherence = decorative writing. High voice + low authority = personality without substance.

1. Coherence Before Surface: Fix logic before fixing rhythm. A polished fallacy is still a fallacy.

1. Authority Is Signaling: We measure expertise markers, not truth. Be honest about this.

1. Hedges Need Classification: Epistemic honesty is not weakness. Only penalize opinion-avoidance.

1. Avoid Overcorrection: "LinkedIn Influencer" mode is not the goal. Calibrate voice to genre.

1. Radical Simplicity Has Value: Sometimes effectiveness means removing signals, not adding them.

Examples

Analysis Example

Input: Business quarterly report with strong metrics but institutional voice

Output:

WRITING EFFECTIVENESS ANALYSIS

Genre: Business | Calibration: Applied

QUALITY PROFILE:
           Craft: ████████░░ 78 - Clean structure, precise numbers
       Coherence: ████████░░ 85 - Logical flow, functional data
       Authority: █████████░ 88 - Genuine SaaS expertise visible
         Purpose: ███████░░░ 75 - Clear reporting, muted stakes
           Voice: ██████░░░░ 62 - Institutional, could be any company
   Effectiveness: ████████░░ 82 - Strong business communication

KEY INSIGHT: High authority through insider metrics (NRR, churn analysis)
compensates for institutional voice. Genre-appropriate execution.

DETAILED ANALYSIS:
...

Elevation Example (Two-Phase)

Phase 1 Output:

STRUCTURAL FIXES:
- Coherence: Moved security section before feature description (foundations first)
- Authority: Replaced "best practices recommend" with specific tradeoff analysis

Phase 2 Output:

STYLISTIC FIXES:
- Rhythm: Added 5-word punch after long explanation
- Commitment: Removed "somewhat" and "fairly" (cowardly hedges)
- Voice: Added one moment of personality without overdoing it

Success Metrics

Objective:

• Coherence score improvement when logic is fixed
• Authority score reflects genuine expertise presence
• No false positives on genre-appropriate conventions
• Dimension gaps correctly diagnose quality issues

Subjective:

• Text reads as effective for its purpose
• Domain experts recognize authentic expertise
• Genre conventions respected, not penalized
• User understands WHY changes improve quality

Notes

• This skill evaluates effectiveness, not truth
• Genre detection happens BEFORE scoring
• Two-phase remediation: structure first, style second
• Hedge classification: epistemic honesty is not weakness
• Avoid overcorrection: "more voice" can become cringe
• Radical simplicity sometimes wins over complexity