script-review — vault-evidence script scoring

Score scripts against what has actually worked for THIS creator, not generic advice. Every claim in the review cites vault data or the creator profile.

Before scoring

1. socials-mcp:profile_get — niche, tone, keywords. If empty, build it first (4 questions + auto-fill).
2. socials-mcp:top_content (limit: 10, metric: "views" and once with metric: "engagement_rate") — the proven patterns.
3. If TikTok retention data exists (vault_query: SELECT title, retention_json, traffic_json FROM video_metrics vm JOIN videos v ON v.id = vm.video_id WHERE retention_json IS NOT NULL ORDER BY captured_at DESC LIMIT 5), read the curves of the top performers.
4. Research search demand — web-search the specific topic, place, product, or person named in the script. Report High / Medium / Low demand and what it implies:
   • High demand (well-known subject): can be a "big swing" piece — expect it to carry reach
   • Low demand (obscure, hyper-local): fine to shoot, often converts viewers into followers well, but raw view count will likely stay modest regardless of quality

Score (0–10 each, weighted total /50)

Good / Bad / Ugly summary

Before the scorecard table, produce three short groups in plain language:

• Good — what's already working, cite the specific line/shot/idea
• Bad — fixable weaknesses, name the fix plainly
• Ugly — anything that risks real underperformance (low-demand topic treated as big-swing, missing keyword entirely, generic title)

Don't manufacture criticism. If a script nails most categories, say so and keep the rewrite section short.

Output format

## Script review — {title}
**Score: {n}/50** — {one-line verdict: produce as-is / revise / rework hook}

### Good / Bad / Ugly
- **Good**: {what works, cite specific line}
- **Bad**: {fixable issue + fix}
- **Ugly**: {real risk, if any}

### Scorecard
| Dimension | Grade | Why | Fix |
|---|---|---|---|
| Hook | {0–10} | {one line} | {concrete fix} |
| Search demand | {0–10} | High/Med/Low + implication | {adjust expectations or lean in} |
| Structure | {0–10} | {one line} | {concrete fix} |
| SEO | {0–10} | {one line} | {concrete fix} |
| Fit | {0–10} | {one line} | {concrete fix} |
| Deliverable safety | {0–10} | {one line} | {concrete fix} |

### Rewrite deliverables (ready to use)
- **Titles**: {2–3, keyword-front-loaded, specific not generic}
- **Caption**: {keyword leading first line, not a vague category tag}
- **Hashtags**: {1 broad + 1 niche + 1 exact-name tag; flag stuffing or zero tags}
- **On-screen text cues**: {hook text at 0s, mid-video callout at ~8–10s, closing CTA}
- **Mid-video re-hook**: {timestamp + what changes on screen or what gets said}
- **Posting time**: {use account data if available, otherwise label as general best-practice}

After review

• Sponsored script → log/stage it: socials-mcp:pipeline_update → script_review, save the script file path.
• Offer to draft the brand review-request email (template in brand-outreach skill assets) and move the item to brand_review when sent.

Rules

• Never invent performance claims ("this will hit 50k") — cite ranges from the creator's own history at most.
• If vault is thin (<5 videos), say so and lean on profile + general craft, marking dimensions with weaker evidence.
• Use plain language — no jargon (CTR, SERP, algorithmic distribution) without immediately explaining it in words a non-technical creator understands.
• The human decides; the review informs.