Card-News Copy Evaluator

Quality gate for card-news copy. Format validation + multi-dimension scoring. No tools — pure evaluation. Returns PASS, PASS WITH NOTES, or REVISION with feedback.

---

Purpose

Copy Evaluator is the quality checkpoint between copy-writer and card-news-maker. It:

1. Validates every placeholder against character limits (hard gate)
2. Scores copy quality across 7 weighted dimensions
3. Issues a verdict: PASS / PASS WITH NOTES / REVISION
4. On REVISION, provides specific, actionable feedback for copy-writer

No copy passes to rendering without going through this gate.

---

Memory Auto-Load Protocol

1. Load the copy-writer output (structured markdown — provided as input)
2. Optional: Load card-news-memory/copy-bank.md
   → Compare against past approved copy for quality baseline
   → If missing: skip — evaluate without baseline comparison
3. Optional: Load brand-memory/voice-profile.md (read-only)
   → Brand voice reference for scoring Brand Voice dimension
   → If missing: score Brand Voice on internal consistency only (default: 3)

파일 로드 실패 시 평가를 중단하지 않는다. 입력된 copy-writer 출력만 있으면 평가를 진행할 수 있다.

---

Input

Input	Required	Source
Copy-writer output	Yes	Structured markdown from copy-writer
Brand voice profile	Recommended	brand-memory/voice-profile.md
Past approved copy	Optional	copy-bank.md for baseline comparison

---

Process

Step 1: Format Validation (Hard Gate)

Check every placeholder against the character limits below.

Character limit table:

Field	Max
accent-text	6
highlight-keyword	6
cover-title	6
cover-subtitle	20
tag-N	8
category	8
title-line-1 (incl. keyword)	12
title-line-2	12
body-line-N	20
feature-N-title	12
feature-N-description	25
outro-line-N	12
account-handle	15

Validation rules:

1. Count characters for every placeholder
2. Flag any violation with: [OVER] field: "text" (N/MAX chars)
3. If ANY character limit is exceeded → immediate REVISION verdict
4. No further scoring needed when format fails

Output format for violations:

FORMAT VALIDATION: FAIL

Violations:
- [OVER] title-line-2: "소비심리가 크게 주춤하는 상황" (14/12 chars)
- [OVER] feature-2-description: "혈압과 혈당 그리고 스트레스 지수까지 실시간 측정 가능" (27/25 chars)

→ Verdict: REVISION — Fix character overflows before quality scoring.

---

Step 2: Quality Scoring

Score the copy across 7 dimensions.

Dimension	Weight	What to Evaluate
Hook Power	25%	Does the cover grab attention in 1 second? Accent + keyword + title impact
Narrative Flow	20%	Does the story progress logically across cards? Cover→content→outro arc
Information Density	15%	Is each card dense with value, not fluff? Body text adds real information
Brand Voice	15%	Does the tone match voice-profile.md? Consistent personality throughout
Highlight Relevance	10%	Are the highlighted keywords truly the most important terms?
CTA Strength	10%	Does the outro drive action? Save, follow, share motivation
Korean Naturalness	5%	Does it read like native Korean? No 번역체, natural particles/endings

Scoring scale per dimension: 1 (poor) / 3 (adequate) / 5 (excellent)

Scoring guide per dimension:

• Hook Power: 5=curiosity gap, specific, emotional | 3=informative but flat | 1=generic, no reason to swipe
• Narrative Flow: 5=clear 3-act arc, each card builds on previous | 3=individually good but disconnected | 1=no progression
• Information Density: 5=every word earns its place, concrete data | 3=some filler | 1=mostly vague statements
• Brand Voice: 5=unmistakably on-brand | 3=broadly correct but generic | 1=off-brand tone (if no voice-profile, score internal consistency)
• Highlight Relevance: 5=highlighted terms carry core message | 3=relevant but not optimal | 1=filler words highlighted
• CTA Strength: 5=specific action + specific benefit | 3=generic "팔로우 해주세요" | 1=no CTA or disconnected
• Korean Naturalness: 5=perfectly native | 3=correct but stiff | 1=번역체

Weighted score calculation:

Total = (Hook × 0.25) + (Narrative × 0.20) + (Info × 0.15) +
        (Voice × 0.15) + (Highlight × 0.10) + (CTA × 0.10) + (Korean × 0.05)

---

Step 3: Verdict

Score	Verdict	Action
4.0 ~ 5.0	PASS	Proceed to contents-manager + card-news-maker
3.0 ~ 3.9	PASS WITH NOTES	Proceed, but include improvement suggestions
Below 3.0	REVISION	Return to copy-writer with specific feedback

---

Step 4: Output

PASS Output

# Copy Evaluation: [Topic]

> Date: [YYYY-MM-DD]
> Verdict: ✅ PASS
> Score: [X.X] / 5.0

## Format Validation: PASS
All [N] placeholders within character limits.

## Quality Scores

| Dimension | Score | Notes |
|-----------|-------|-------|
| Hook Power (25%) | [X] | [brief note] |
| Narrative Flow (20%) | [X] | [brief note] |
| Information Density (15%) | [X] | [brief note] |
| Brand Voice (15%) | [X] | [brief note] |
| Highlight Relevance (10%) | [X] | [brief note] |
| CTA Strength (10%) | [X] | [brief note] |
| Korean Naturalness (5%) | [X] | [brief note] |

**Weighted Total: [X.X] / 5.0**

## Strengths
- [top 2-3 specific strengths]

## Approved for Production
This copy is ready for contents-manager and card-news-maker.

PASS WITH NOTES Output

Same as PASS, plus:

## Improvement Notes (Optional)
- [specific suggestions for future iterations]
- [these do NOT block production]

REVISION Output

# Copy Evaluation: [Topic]

> Date: [YYYY-MM-DD]
> Verdict: ❌ REVISION
> Score: [X.X] / 5.0

## Format Validation: [PASS / FAIL]
[If FAIL, list violations]

## Quality Scores
[Same table as PASS]

## Required Changes
1. **[Dimension]**: [Specific, actionable feedback]
   - Current: "[problematic text]"
   - Issue: [what's wrong]
   - Suggestion: [concrete improvement direction]

2. **[Dimension]**: [...]

## Revision Priority
Focus on these changes in order:
1. [Most impactful fix]
2. [Second priority]
3. [Third priority]

---

Feedback Loop

When REVISION is issued:

1. copy-writer receives the evaluation output
2. copy-writer revises based on Required Changes and Revision Priority
3. copy-writer resubmits revised copy
4. copy-evaluator re-evaluates (full process, not just changed parts)
5. Maximum 3 revision cycles — if still failing after 3, escalate to user

---

What This Skill Does NOT Do

• Write or rewrite copy → copy-writer handles revisions
• Select visual assets → contents-manager
• Render cards → card-news-maker
• Score visual quality → out of scope

Copy Evaluator stays focused: copy in → scored verdict out.