UGC Script Writing

Write scripts people actually want to watch. Every script is built for someone talking to a camera — a real person or an avatar. The goal is a script that sounds natural when spoken aloud, holds attention through the first 3 seconds, and delivers one clear idea.

Outcome

Platform-ready scripts saved to projects/mkt-ugc-scripts/{YYYY-MM-DD}_{batch-name}/. Each script includes timing cues, delivery notes, on-screen text suggestions, and a hook designed for talking-head delivery.

Context Needs

Load what exists. Skip what doesn't. This skill works standalone.

Dependencies

Skill Relationships

• Upstream: str-trending-research (topic angles), mkt-brand-voice (voice context), mkt-icp (audience language)
• Downstream: viz-ugc-heygen (takes finished scripts and generates avatar videos), mkt-content-repurposing (scripts can be atomized into text posts)

Step 1: Load Context

Check brand_context/ and load per the table above. Show a brief status:

Script context:
  Voice Profile     [loaded / not found]
  Positioning       [loaded / not found]
  ICP               [loaded / not found]
  Learnings         [N entries / none yet]

Read context/learnings.md → ## mkt-ugc-scripts. Apply any previous corrections before writing.

If voice profile is missing, work with a natural conversational tone and mention that /mkt-brand-voice would sharpen the scripts.

Step 2: Topic Selection

Two paths:

Path A — User has a topic: They tell you what to script. Confirm the core message and move to Step 3.

Path B — Research-driven: If the user wants ideas, or says something like "what should I make a video about":

1. Ask: "What's your niche or business area? I'll research what's trending."
2. Run str-trending-research on their niche (or read a recent research brief if one exists in projects/str-trending-research/)
3. Read references/research-to-script.md for the methodology on turning research into script angles
4. Present 3-5 script-worthy angles with a one-line pitch for each
5. User picks one (or combines)

Either path: confirm one core idea before writing. A script that tries to say two things says nothing.

Step 3: Pick Script Format + Platform

Ask (or infer from context):

• Platform target: TikTok, Instagram Reels, YouTube Shorts, or multi-platform
• Duration target: 15s, 30s, 45s, 60s, 90s max (default: 30s)

Then read references/script-frameworks.md and pick the format that fits the topic:

Tell the user which format you're using and why. If they want a different one, switch.

Step 4: Write the Script

Read references/script-frameworks.md for the chosen format's structure and references/hook-library.md for hook options. Read assets/reference-script-openclaw-loop.md as an illustration of the personal-experience tone: specific details, natural flow, no listicle structure.

The script IS the spoken words. Nothing else. No timestamps like [0-3s], no stage directions like [lean in], no [ON SCREEN:] annotations. The output must be directly pasteable into HeyGen or readable as a teleprompter script. Follow the scripting conventions from viz-ugc-heygen/references/scripting-guide.md.

Script format:

PLATFORM: [TikTok / Reels / Shorts]
DURATION: [estimated seconds]
FORMAT: [framework name]
WORD COUNT: [number]

---

The exact spoken words go here. Every sentence is what the person says out loud.

Add a break tag only where a dramatic beat needs one.

Continue with the next section. No labels, no annotations, no brackets.

The CTA closes the script naturally.

On-screen text suggestions go in a separate section AFTER the script body:

ON-SCREEN TEXT:
- [when]: [text overlay]
- [when]: [text overlay]

This keeps the script clean and pasteable while still providing production notes.

Writing rules for spoken delivery:

• 150 words per minute at natural pace — use this to estimate duration
• Keep each sentence short enough to say in one breath; split stacked clauses, since TTS voices flatten them.
• Use contractions. "It's" not "it is". "Don't" not "do not".
• Write how people talk, not how they write. Read it aloud.
• No parenthetical asides — AI voices struggle with them
• Front-load the hook — the first sentence decides if they stay
• One idea per beat. Don't stack concepts.
• Add <break time="1.2s"/> (1.2-1.5s) only where a dramatic beat needs it; most scripts need none.

Present the script to the user. This is a creative checkpoint — wait for their feedback before finalising.

Step 5: Batch Mode (If Asked)

If the user wants multiple scripts:

• Write 3-5 scripts per batch
• Vary the formats across the batch (don't write 5 quick teaches)
• Each script gets the same full treatment from Step 4
• Present all scripts together for review

Step 6: Humanizer Gate

Before saving, run all scripts through tool-humanizer in pipeline mode. Use deep mode if brand_context/voice-profile.md was loaded, standard otherwise. Only show the score summary if the delta exceeds 2 points.

Spoken scripts should score high naturally since they're already conversational — if the humanizer is changing a lot, the script probably needs a rewrite, not just pattern removal.

Step 7: Save Output

Save every script to disk.

Create folder: projects/mkt-ugc-scripts/{YYYY-MM-DD}_{batch-name}/

One file per script: {topic-slug}_{platform}.md

Each file contains:

1. YAML frontmatter with metadata
2. The script body (spoken words only)
3. ON-SCREEN TEXT section after a --- divider

---
platform: tiktok
format: personal-experience
duration_estimate: 45s
word_count: 112
topic: "the core topic"
date: YYYY-MM-DD
---

The spoken script goes here exactly as written...

---

ON-SCREEN TEXT:
- Hook: text overlay
- Mid: text overlay
- Close: text overlay

After saving, show actual script excerpts — not just file paths.

Suggest next steps:

• "Want to generate this as a video? I can send it to HeyGen." (→ viz-ugc-heygen)
• "Want more scripts on related angles?"
• "Want to repurpose this script into text posts?" (→ mkt-content-repurposing)

Step 8: Collect Feedback

Ask: "How do these read? Anything feel off when you imagine saying it out loud?"

Log feedback to context/learnings.md → ## mkt-ugc-scripts with date and context.

Rules

• 2026-03-10: Maximum script duration is 90 seconds. Never write a script longer than 90s. Most scripts should land between 15-60s.
• 2026-03-10: End every script with a soft, conditional CTA pointing to the community community link in bio, phrased differently for each script in a batch (for example, "If you want the full breakdown, the free community link is in my bio").
• 2026-03-10: Scripts must be the exact spoken words only — no timestamps like [0-3s], no stage directions like [lean in], no [ON SCREEN:] inline. Output must be directly pasteable into HeyGen or usable as a teleprompter. Use SSML <break time="Xs"/> for pauses only where a dramatic beat is needed — don't litter the script with them. Most scripts flow naturally without any. On-screen text suggestions go in a separate section after the script body.
• 2026-03-10: Scripts must sound like personal experience, not teaching or selling. Write in first person about what YOU did, what YOU switched to, what problems YOU had. Name specific tools, specific frustrations, specific results. The audience learns by overhearing your experience, not by being lectured. "I will never use X again because..." beats "Stop using X. Here's why." No promotional language, no listicle energy, no "here are 3 tips" structure unless specifically asked for. The CTA should be soft and conditional: "If you want the full breakdown, link's in my bio" not "Join the free community!"

Self-Update

If the user flags an issue with the output, add a dated entry under ## Rules in this skill's SKILL.local.md (create it if missing) and log it to context/learnings.md under ## mkt-ugc-scripts. Write the general rule the correction implies, not the one-off fix.