Write Script

Step 2: Research

Run 2-3 web searches to gather:

• Current facts, stats, and status of the topic
• What people are saying about it right now
• Any recent developments the script should reference
• Traceable source URLs to include in the output

Step 3: Retrieve voice context

Run: cd pipelines && python retrieve_similar.py "CORE TOPIC" -n 3 Read the output to see how Niyi has written about similar subjects before.

Step 4: Generate 2-3 script variations

All variations use the same angle (the user's stance) but with different structures:

Variation 1: Rant/Story structure Hook → Personal take → Build the argument with facts → Reality check → CTA

Variation 2: List/Steps structure Hook → Numbered points (3-5) → Standout insight → CTA

Variation 3: Question/Answer structure Hook (posed as a question) → Explore the answer → What most people miss → CTA

For each variation, generate:

• 3 hook options (each using a DIFFERENT pattern from the hook patterns file — rotate through Bold Claim, Imagine, What If, Did You Know, If There Is, Direct Challenge, Personal Story, List Intro, Problem/Pain Point, If You Qualifier)
• Full script (~150-200 words)
• Caption (social media caption with hashtags)
• 3 comment prompts (engagement questions)
• Sources (traceable URLs from web search, 2-4 per variation)

Key Rules

1. Preserve the user's exact phrasing. If they said something with energy, personality, or strong emotion — keep it word for word. Build the polished script AROUND their raw voice, don't replace it.

2. Voice matching checklist (apply to EVERY variation):

   • Does it sound conversational, not formal?
   • Does it use "you" and "your" directly?
   • Does it include a practical takeaway?
   • Does it use analogies to explain complex ideas?
   • Would Niyi actually say this out loud?
   • Is it 150-200 words?

3. Structure rules by content type:

   • HOW_TO: Hook → Problem → Steps (3-5) → Result → CTA
   • OPINION: Hook → Take → Evidence → Reality check → CTA
   • ROUNDUP: Hook → Numbered list → Standout pick → CTA
   • NEWS: Hook → What happened → Why it matters → What's next → CTA

Step 5: Save outputs

Create output directory if needed: mkdir -p outputs/custom

Save two files:

• outputs/custom/YYYY-MM-DD_[slug].json — Full script data (see schema below)
• outputs/custom/YYYY-MM-DD_[slug].md — Human-readable markdown

The slug should be a short kebab-case version of the topic (e.g., chatgpt-atlas, vibe-coding-beginners).

Step 6: Present results

Show all variations side by side. Note which phrases were preserved from the user's original input.

Output JSON Schema

{
  "generated_at": "ISO8601 timestamp",
  "input": "the raw user input (full text)",
  "topic": "extracted core topic",
  "content_type": "OPINION|HOW_TO|ROUNDUP|NEWS",
  "angle": "the detected angle/stance",
  "preserved_phrases": ["phrase 1 kept verbatim", "phrase 2 kept verbatim"],
  "research_summary": "key facts found during web search",
  "variations": [
    {
      "id": 1,
      "structure": "rant|list|question",
      "hooks": ["hook 1", "hook 2", "hook 3"],
      "script": "full 150-200 word script",
      "caption": "social media caption with hashtags",
      "comment_prompts": ["question 1", "question 2", "question 3"],
      "sources": ["https://...", "https://..."]
    }
  ]
}

Output Markdown Format

# Custom Script: [Topic]

> Generated from: "[first 100 chars of input]..."
> Type: [OPINION] | Angle: [CONTRARIAN/FRUSTRATION/etc.]
> Preserved phrases: "[phrase 1]", "[phrase 2]"

## Research Summary
[Key facts, stats, and current status found during web search]

---

## Variation 1: Rant/Story
### Hooks
1. ...
2. ...
3. ...

### Script
[Full script]

### Caption
[Caption with hashtags]

### Comment Prompts
- ...

### Sources
- https://...

---

## Variation 2: List/Steps
[Same format]

---

## Variation 3: Question/Answer
[Same format]