UGC Ad Strategist — the research + scriptwriting brain skill

UGC Ad Strategist — the research + scriptwriting brain

ugc-video-producer can make a real-looking person say words with synced lips and show a real product. It has no brain for the words. This skill IS that brain: market research → one claim → killer-script copy → brief.json. The producer consumes brief.json and renders the video.

Scripting engine: this skill delegates the narrative craft to the killer-script-system skill (installed active at ~/.claude/skills/killer-script-system/) — Packaging/Outline pick the claim + angle, Intro/Body/Outro + native-CTA write the copy. This skill wraps it with the two things it lacks for paid UGC: live research and short-form ad shaping.

MINIMAX MODE — offload the WRITING to Claude (self-test: can't see an image = MiniMax). MiniMax writes stiff, ad-flavored dialogue. When on MiniMax this skill is Step 2 of the ugc-minimax-mode pipeline (follow that skill for the whole job). Do NOT let the Script stage write the copy itself — run ~/.openclaw/scripts/ugc_write_script.sh --product "<...>" --context "<...>" --format testimonial|two_person_interview|street_interview --clips <n> --out <run>/script.json (natural, brief-schema JSON), then merge into brief.json. Research + brief stages can stay on MiniMax; only the dialogue/hook WRITING needs Claude. On Codex/gpt-5.5, skip this — Codex writes natural copy itself.

Generation principles: the script you produce gets rendered by ugc-video-producer, whose reference/ugc_qa_bank.md is the bank of real production mistakes-turned-rules. The script-side ones (A1 size dialogue to 6/10s clips, A2 only script REAL on-screen assets, A3 truthful scarcity, A4 the non-negotiables) are enforced in the Script stage below — honor them.

When to run

• The user asks what an ad should say / to write ad copy / to research the angle for a product.
• ugc-video-producer Phase 0 hands off a product (not a finished script).
• Fully autonomous by design: research → finished brief.json with no stops.

How to run it (the launch sequence)

It's a Workflow-script pipeline (deterministic research fan-out), same pattern as ebook-from-scratch and deep-research. The Workflow args channel is unreliable on this machine (see the ebook args-not-delivered note), so the product spec is passed via a request.json file the workflow reads, not via args alone.

1. Write the request. Pick a slug, make a run dir, write request.json:

   SLUG="aimastery-$(date +%s)"; RUN=~/aimastery_ads/strategist_runs/$SLUG; mkdir -p "$RUN"
   cat > "$RUN/request.json" <<'JSON'
   { "product": "The Complete AI Mastery Collection",
     "company": "claritybooks",
     "kind": "digital",
     "country": "US",
     "voice_id": "kPzsL2i3teMYv0FxEYQ6",
     "competitors": [],
     "notes": "4 books, $59.99 one-time. LP: aimastery-salespage.vercel.app" }
   JSON
   

   Only product is required; the workflow resolves the rest (company → ~/.openclaw/companies.json, funnel → the live landing page, assets → the real PDFs).

2. Launch the workflow, passing the request path via args AND relying on the newest-request fallback: Call the Workflow tool with { scriptPath: "~/.claude/skills/ugc-ad-strategist/scripts/strategist_workflow.js", args: { requestPath: "<RUN>/request.json" } }.

3. On return, the workflow gives you brief_path (default <RUN>/brief.json), the single_claim, clip_count, and hook_count. Read brief.json and hand it to ugc-video-producer.

The 5 stages (what the workflow does)

Stage	Job
A · Ground	product → company → LIVE funnel (offer, price, anchor, exact LP headlines for message-match), ad industry, niche keywords, real on-screen PDF assets.
B · Research	3 parallel streams: voice-of-customer (~/code/kojo-audience-scraper/exports/), market/angle intel (web research: winning angles + 2026 scam-flag claim patterns to avoid + awareness stage + sophistication), proven-winner ad mining (scripts/ad_library_scout.py → Meta Ad Library + TikTok Creative Center, ranked by longevity).
C · Brief	killer-script Packaging+Outline → ONE claim (capability, never income), angle, proof object, objections, message-match line, differentiation vs in-market winners.
D · Script	killer-script Intro/Body/Outro → 6 hook variants, the 6-beat sheet (hook→proof→mechanism→escalation→honest-80/20→CTA) with per-beat dialogue/emotion/action/grok_prompt/overlay, platform cuts (paid speaks price for FTC; organic DM CTA), voice direction, UTM map.
E · Assemble	Merge everything into brief.json (the producer contract).

Output contract — brief.json

Defined by reference/brief_schema.json. The producer already expects per-clip dialogue, base_still_prompt, voice.id, and (digital) screen_asset.pages. Key fields: single_claim, clips[], hooks[], platform_cuts, voice, utm_map, funnel, market_winners[].

The non-negotiables (baked into the template, enforce them)

One claim per ad · capability not income · first-person past-tense testimonial · process proof beats outcome proof · honest 80/20 to pre-handle the objection · persistent "results vary" disclaimer spoken + on-screen · price spoken on paid · caption every beat. These are house-style writing rules — write to them as you draft, not as a fix after the fact. Full rationale + the killer-script→UGC beat map: reference/ad_package_template.md.

Tools / standalone use

• Ad mining only: python3 scripts/ad_library_scout.py --query "AI course" --min-days 30 --json. Degrades to a URLs-to-check result when Meta is bot-gated — never fabricates ads.

Gotchas

• Reuses, doesn't reinvent: killer-script-system (copy), llm.sh, the kojo exports, deep-research (heavier fallback for Stage B2), ugc-device-screen. Don't duplicate these.

Companion skills

• ugc-video-producer — consumes brief.json, renders the video (this skill is its Phase 0).
• killer-script-system — the scripting engine this skill drives.
• ugc-device-screen — pins the real PDF on the device for screen_asset proof clips.