Reels Production Agent

A Claude Code agent that generates production-ready Instagram Reels / YouTube Shorts / TikTok content batches.

What it produces per batch:

• 30 teleprompter-ready scripts (word-for-word, just read and record)
• 30 deliverables (the actual prompts, templates, and guides promised in CTAs)
• Production checklist (recording prep, screen recording setup, ManyChat keywords)
• Batch history tracking (cross-batch deduplication so content never repeats)

How It Works

Three specialized Claude subagents run in sequence:

1. Researcher (Sonnet) — Discovers and verifies 30-35 AI tools via web search. Confirms pricing from official pages. Captures URLs, competitors, and demo access.
2. Scriptwriter (Opus) — Writes 30 scripts from the research data using strict voice calibration, 14 hook formulas, 4 templates, and quality rules derived from real high-performing Reels.
3. Deliverable Builder (Sonnet) — Reads finished scripts and builds the actual assets promised in CTAs: prompt packs, setup guides, workflow templates.

Requirements

• Claude Code CLI installed
• Claude API key with access to Opus and Sonnet models

Quick Start

cd ReelsProduction

# Generate a test batch (5 scripts) to check quality
claude /generate-reels-test

# Generate a full batch (30 scripts)
claude /generate-reels

Customizing for Your Niche

This agent is built for AI tools for e-commerce brands, but the architecture works for any niche:

1. Update brand context in CLAUDE.md — your niche, positioning, and creator bio
2. Update strategy in reference/strategy.md — your TAM filters, hook examples, and topic selection criteria
3. Update research scope in .claude/agents/researcher.md — your 5 sub-themes and what to search for
4. Update voice in .claude/agents/scriptwriter.md — your reference transcripts (paste 3 of your best-performing scripts as the quality bar)

The deliverable builder and slash commands work unchanged for any niche.

Project Structure

ReelsProduction/
├── CLAUDE.md                      # Pipeline orchestration instructions
├── .claude/
│   ├── agents/
│   │   ├── researcher.md          # Tool discovery & verification
│   │   ├── scriptwriter.md        # Script writing with voice calibration
│   │   └── deliverable-builder.md # CTA deliverable creation
│   └── commands/
│       ├── generate-reels.md      # /generate-reels (30 scripts)
│       └── generate-reels-test.md # /generate-reels-test (5 scripts)
├── reference/
│   └── strategy.md                # Strategic thinking layer (TAM, TOFU, hooks)
└── output/                        # Generated batches (gitignored)
    ├── batch_history.md           # Cross-batch deduplication log
    └── batch_YYYY-MM-DD/          # One folder per batch

Architecture Decisions

• Three separate subagents instead of one monolith: each stays focused, avoids context window bloat, and can be swapped independently.
• Scriptwriter never researches: it only writes from verified data. This prevents hallucinated tool claims.
• Researcher never writes scripts: it only verifies tools exist and captures accurate data.
• Batch history as a flat file: simple, portable, and readable by all subagents without tooling.
• 3 groups of 10: the scriptwriter generates in batches of 10 to prevent quality degradation as the context window fills.

License

MIT