Script Pipeline Skill

Purpose

End-to-end script generation that ensures every video has:

1. A unique, researched topic (never repeated on the same channel)
2. A hook optimized for 30-second retention (using ranked formulas)
3. A structure that maximizes watch time (micro-loops, pattern interrupts, strategic silence)

When to Use

• Before rendering any video (this is the first step in the pipeline)
• When the user asks to "generate a script" or "make a video for channel X"
• As part of the automated pipeline (scheduled or manual trigger)

Prerequisites

• config/channels.json must exist with channel configurations
• data/topic-log.json must exist (deduplication tracking)
• data/hook-templates.json must exist (hook formulas)
• Internet access for web research

Process

Phase 1: Topic Research

1. Read config/channels.json to get channel niche, style, and tone
2. Read data/topic-log.json to get previously used topics for this channel
3. Run 3-4 search passes per the deep-research skill:
   • Pass 1: Broad web search on the niche
   • Pass 2: Narrow/factual pass with specific terms
   • Pass 3: Practitioner pass (what other YouTube channels in this niche are covering)
   • Pass 4: Social/community pass (Reddit, forums, trending discussions)
4. Filter out any topics that overlap with used_topics in the topic log
5. Rank remaining topics by: trending potential, freshness, RPM alignment, hook potential

Phase 2: Topic Selection & Deduplication

1. From researched topics, select the top candidate
2. Check against data/topic-log.json for semantic duplicates (not just exact matches) — use the code-level check:
   • node src/utils/topic-log.cjs <channel-id> list — show every already-used topic
   • node src/utils/topic-log.cjs <channel-id> pick "<candidate>" — exits non-zero if the candidate is a duplicate
   • Also honor data/research/<channel-id>/next-topic.json if it exists (the pipeline already reserved that topic as the next one)
3. If topic is too similar to an existing one, reject and try next candidate
4. Once confirmed unique, reserve it: node src/utils/topic-log.cjs <channel-id> reserve "<topic>". script-writer/run.js does this automatically every time it writes a script — never skip that step.

Phase 3: Hook Optimization

1. Read data/hook-templates.json → niche_hook_mapping for this channel ID
2. Select the primary hook type for this niche
3. Generate 3 hook variations using the ranked formulas
4. Pick the hook with highest expected retention (specific-result > contrarian > direct-promise)
5. Ensure hook follows rules:
   • Under 15 seconds spoken
   • Opens with number, name, or strong action verb
   • No "welcome to" or channel name
   • No questions as openers (make statements)

Phase 4: Script Generation

1. Read channel config for style, tone, pacing
2. Load the appropriate template (cinematic-documentary, minimal, or motion-graphics)
3. Generate script following the script-writer skill's workflow
4. Apply retention techniques:
   • Micro-loops at end of each section
   • Pattern interrupts every 60-90 seconds
   • Strategic silence after reveals (750-1500ms)
   • Callback close to hook
   • Statistics with context
5. Output the full script JSON per the schema

Phase 5: Update Tracking

1. Add the topic to data/topic-log.json under the channel's used_topics
2. Update last_generated timestamp
3. Save the script to data/scripts/{channel_id}/{topic-slug}.json

Output

• Script JSON file at data/scripts/{channel_id}/{topic-slug}.json
• Updated data/topic-log.json with the new topic logged
• Summary to user with: topic, hook, word count, estimated duration, sections

Rules

• Never generate a topic that's already in the channel's used_topics list (check via node src/utils/topic-log.cjs <channel-id> list)
• Every script written must be registered in data/topic-log.json (script-writer does this; verify with node src/utils/topic-log.cjs <channel-id> list)
• Every fact in the script must come from actual web research (not general knowledge)
• Hook must be from the ranked formulas — don't invent new hook styles
• Follow the channel's style template exactly (pacing, transitions, SFX)
• If research doesn't support a needed claim, flag the gap — don't invent