SuperAPI AI Directory

scrollclaw-persona

General & other

View repository

Updated 10-16 07:15

About this skill

Persona & Script

Everything starts here. No first frames, no animation, no audio until the messaging is locked.

Step 1: Persona Research (do not skip)

Read references/persona-research.md. Minimum 60 minutes mining competitor reviews on Amazon, Trustpilot, Reddit. Copy exact phrases — do not paraphrase.

You're looking for exact language, not information:

  • "I was doing everything right and my body wasn't cooperating" → goes into a script
  • "I finally feel like myself again" → useless, discard
  • "I wore a dress I hadn't put on since my daughter was born and cried in the fitting room" → this is what you want

Specific language produces viewers who can viscerally imagine the result. Generic language produces viewers who merely understand it's possible. The first converts. The second doesn't.

Step 2: Brand Context

Load brand voice file if it exists. Know the product, the audience, the vibe.

Step 3: Creator Profiles

Each creator gets a profile in workspace/campaigns/<slug>/creators/creator-<name>.md for campaign-specific work, or workspace/creators/creator-<name>.md for reusable global profiles. Read references/script-voice.md for voice calibration.

bash scripts/create-creator.sh <workspace> <creator-name>

Step 4: Format Selection

Refer to the format table in the system context (loaded automatically). Ask the user what they're making and guide them.

For Hook Face + Demo: always recommend real demo footage (screen recording) over AI-generated demo. The hook face stops the scroll — the demo should be the real product. Use scripts/build-hook-demo.sh to combine.

Step 5: Script

Read references/script-voice.md. Use exact phrases from persona research. Test every line: if it sounds like a copywriter wrote it, rewrite until it sounds like someone talking to their phone.

FORMAT ENFORCEMENT (mandatory): Before writing, pull the shot breakdown from the format table in the system context (loaded automatically) for the chosen format. The script MUST map to the shot breakdown:

Format Script must have
Talking Head Hook (0-4s) + Show (4-12s) + Verdict (12-20s). SHOW must include a visual action.
Hook Face + Demo Hook face with emotion + text (0-4s) + hard cut + Demo (4-14s). Hook has NO dialogue.
Podcast Clip Mid-conversation entry + The take + The beat. Must reference unseen host.
Visual Transformation Hook frame with concept name + Before + Turning point + After + Close
Hybrid Transformation Talking head before + Slideshow mechanism bridge + Talking head after
Wall of Text Dense text only — no voice, no video gen

Validation: After writing, map each line to a segment. If any segment is missing or the script is one continuous monologue with no visual structure — rewrite. A script without visual beats is not a script, it's a ramble.

Present the script to the user for approval before generating anything. Include the segment mapping. User can override with "skip approval" for autonomous generation.

Mark the script with [A-ROLL] and [B-ROLL] tags. A-roll = Sora (talking head + dialogue). B-roll = Kling (everything else). Voice runs continuously over B-roll.

Brand Memory Integration

Reads

File Purpose
workspace/brand/voice-profile.md Script tone calibration — if brand voice is "conversational-direct," lean into informal breaks and real pauses
workspace/brand/positioning.md Which pain points to emphasize; how the product differentiates
workspace/brand/audience.md Creator archetype selection — anchors age, lifestyle, aesthetic
workspace/campaigns/<slug>/brief.md Campaign-specific product, goal, and format direction

Writes

File Notes
workspace/campaigns/<slug>/persona-research.md Extracted phrases, pain points, exact customer language
workspace/campaigns/<slug>/creators/creator-<name>.md Campaign-specific creator profile
workspace/creators/creator-<name>.md Global reusable profile (when creator will appear across campaigns)
workspace/campaigns/<slug>/scripts/<format>-script.md Approved script with A/B-roll tags

Context loading (show this to the user at session start)

📋 Context loaded for campaign: <slug>
  ✓ Brand voice: <tone from voice-profile.md> (workspace/brand/voice-profile.md)
  ✓ Positioning: <one-line summary> (workspace/brand/positioning.md)
  ✗ No audience file found — creator archetype will be inferred from campaign brief
  ✓ Campaign brief: workspace/campaigns/<slug>/brief.md

Handle missing files gracefully. Never error. Proceed standalone with a note.

Contract

Input

  • Required: campaign brief plus a clear product, audience, or offer to write from
  • Optional: workspace/brand/{voice-profile,positioning,audience}.md, existing creator profiles, format preference
  • Format: brief, raw brand notes, review mining inputs, and workspace markdown files
  • Source: user prompt, workspace/campaigns/<slug>/brief.md, and upstream brand memory files

Output

  • Produces: persona research, creator profile(s), and one approved script with [A-ROLL] and [B-ROLL] tags
  • Format: markdown files in workspace/campaigns/<slug>/ plus inline script review for approval
  • Default behavior: do the research, choose the format, and present the script for approval before any visual generation
  • Downstream use: /first-frame, /animate, /b-roll, /assemble, /score

Validation

  • Pre-conditions: campaign brief exists and there is enough product context to identify a real customer problem
  • Post-conditions: script maps to the format blueprint, uses exact customer language, and has explicit visual beats
  • Failure checks: do not advance with generic copy, missing segment mapping, or an unapproved script unless the user explicitly says to skip approval

Output

  • Persona research doc with extracted phrases — workspace/campaigns/<slug>/persona-research.md
  • Creator profile(s) in workspace/campaigns/<slug>/creators/ (or workspace/creators/ for global)
  • Approved script with segment mapping and A/B-roll tags — workspace/campaigns/<slug>/scripts/<format>-script.md
  • Format selection locked

Next Step

Script approved → run /first-frame to generate the canonical face image.