专家评审 Skill

Preamble (runs on skill start)

# Version check (silent if up to date)
python3 telemetry/version_check.py 2>/dev/null || true

# Telemetry opt-in (first run only, then remembers your choice)
python3 telemetry/telemetry_init.py 2>/dev/null || true

Privacy: This skill logs usage locally to ~/.ai-marketing-skills/analytics/. Remote telemetry is opt-in only. No code, file paths, or repo content is ever collected. See telemetry/README.md.

---

Expert Panel

General-purpose scoring and iterative improvement engine. Auto-assembles the right experts for whatever is being evaluated, scores it, and loops until 90+.

---

Step 1: Intake — Understand What's Being Scored

Collect or infer from context:

1. Content/artifact — The thing(s) to score (paste, file path, or URL)
2. Content type — Copy, sequence, landing page, strategy, title, chart, candidate eval, etc.
3. Offer context — What's being sold/promoted? To whom? What domain/industry?
4. Variants — Are there multiple versions to compare? (A/B/C)
5. Source skill — Is this output from another skill? (e.g., cold-outbound-optimizer) If yes, note the source for feedback-to-source routing in Step 6.

If context is obvious from the conversation, don't ask — just proceed.

---

Step 2: Auto-Assemble the Expert Panel

Build a panel of 7–10 experts tailored to the content type and domain.

Assembly rules

1. Start with content-type experts. Read experts/ directory for pre-built panels matching the content type. If an exact match exists (e.g., experts/linkedin.md for a LinkedIn post), use it as the base.

2. Add domain/offer experts. Based on the offer context, add 1–3 experts who understand the specific industry or domain. Examples:

   • Scoring bakery marketing → add Food & Beverage Marketing Expert
   • Scoring SaaS landing page → add SaaS Conversion Expert
   • Scoring recruiting outreach → add Agency Recruiter + Talent Market Expert
   • Scoring medical device copy → add Healthcare Compliance Expert

3. Always include these two:

   • AI Writing Detector — See experts/humanizer.md. Weight: 1.5x. Non-negotiable.
   • Brand Voice Match — Checks alignment with the configured brand voice and known rejection patterns from references/patterns.md (if present).

4. Check learned patterns. If references/patterns.md exists, read it. If any patterns apply to this content type, brief the panel on them. Dock points for known-bad patterns.

5. Cap at 10 experts. If you have more than 10, merge overlapping roles.

Panel output format

List each expert with: Name, lens/focus, what they check.

---

Step 3: Select Scoring Rubric

Choose the appropriate rubric from scoring-rubrics/:

Content type	Rubric file
Blog, social, email, newsletter, scripts	scoring-rubrics/content-quality.md
Strategy, recommendations, analysis	scoring-rubrics/strategic-quality.md
Landing pages, ads, CTAs	scoring-rubrics/conversion-quality.md
Charts, data viz, infographics	scoring-rubrics/visual-quality.md
Candidate evaluations	scoring-rubrics/evaluation-quality.md
Other	Synthesize a rubric from the two closest matches

Read the selected rubric file for detailed criteria and point allocation.

---

Step 4: Score — Recursive Loop Until 90+

Target: 90/100 across all experts. Non-negotiable. Max 3 rounds.

Each round produces:

## Round [N] — Score: [AVG]/100

| Expert | Score | Key Feedback |
|--------|-------|--------------|
| [Name] | [0-100] | [One-line rationale] |
| ... | ... | ... |

**Aggregate:** [weighted average — humanizer at 1.5x]
**Top 3 weaknesses:** [ranked]
**Changes made:** [specific edits addressing each weakness]

Then the revised content/artifact.

Rules

• Scores must be brutally honest. No padding to 90.
• Humanizer score weighted 1.5x in the aggregate.
• If aggregate < 90: identify top 3 weaknesses → revise → next round.
• If aggregate ≥ 90: finalize and proceed to output.
• After 3 rounds, if still < 90: return best version with honest score + note on what's holding it back.
• Show ALL rounds in output — the iteration trail is part of the value.

Variant comparison mode

When scoring multiple variants (A/B/C):

• Score each variant independently through the full panel.
• After scoring, rank variants by aggregate score.
• If top variant is < 90, iterate on the best one (don't iterate all of them).

---

Step 5: Output Format

Winner + Score (always at top)

## 🏆 Result: [SCORE]/100 — [PASS ✅ | NEEDS WORK ⚠️]

[Final content/artifact here]

**Iterations:** [N] rounds
**Panel:** [Expert names, comma-separated]

If variants: show winner first, then runner-up scores.

## 🏆 Winner: Variant [X] — [SCORE]/100

[Winning content]

### Runner-up scores
- Variant A: 87/100
- Variant B: 82/100
- Variant C: 91/100 ← Winner

Feedback History (below the result)

Show full scoring rounds.

---
<details>
<summary>📊 Scoring History (N rounds)</summary>

[All round tables from Step 4]

</details>

---

Step 6: Feedback-to-Source (When Scoring Another Skill's Output)

When the scored content came from another skill, generate a Source Improvement Brief:

## 🔁 Feedback for [Source Skill]

### What scored low
- [Pattern]: [Specific example from this content]

### Suggested skill improvements
- [Concrete change to the source skill's process/rubric/prompt]

### Patterns to add to source skill
- [Any recurring weakness that should become a rule]

This brief can be used to update the source skill's SKILL.md or rubrics.

---

Step 7: Memory — Learn from Approvals and Rejections

After the user approves or rejects panel output:

On approval (score ≥ 90, user accepts)

Note what worked. No action needed unless a new positive pattern emerges.

On rejection (user overrides the panel or rejects 90+ content)

1. Ask why (or infer from context).
2. Add a new pattern to references/patterns.md using this format:

## [Pattern Name]
- **Type:** rejection | preference | override
- **Content types:** [which types this applies to]
- **Rule:** [What to always/never do]
- **Example:** [The specific instance that triggered this]
- **Date:** [YYYY-MM-DD]
- **Point dock:** [-N points when detected]

3. Confirm: "Added pattern: [one-line summary]. Panel will dock [N] points for this going forward."

Pattern enforcement

Every scoring round, check references/patterns.md against the content. Apply point docks before expert scoring begins. This means known-bad patterns are penalized even if individual experts miss them.

---

Reference Files

File	Purpose	When to read
experts/humanizer.md	AI writing detection rubric (24 patterns)	Every scoring run
experts/[domain].md	Pre-built expert panels for common domains	When domain matches
scoring-rubrics/content-quality.md	Content scoring rubric	Content scoring
scoring-rubrics/strategic-quality.md	Strategy scoring rubric	Strategy scoring
scoring-rubrics/conversion-quality.md	Landing page/ad/CTA rubric	Conversion scoring
scoring-rubrics/visual-quality.md	Chart/data viz/infographic rubric	Visual scoring
scoring-rubrics/evaluation-quality.md	Candidate/assessment rubric	Eval scoring
references/patterns.md	Learned rejection patterns	Every scoring run
references/expert-assembly.md	Domain-expert examples for auto-assembly	When building unfamiliar panels