# Director Structured Captioning

> STRUCTURED CAPTIONING — separate global context from per-shot details (HoloCine pattern):

- Skill: `2799662352/director-structured-captioning` (Agent Skill)
- Install (CLI): `npx skillmds@latest add 2799662352/director-structured-captioning`
- Raw SKILL.md: https://api.skillmd.com/api/skills/2799662352/director-structured-captioning/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: 2799662352 (https://skillmd.com/u/2799662352)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/2799662352/director-structured-captioning

---


STRUCTURED CAPTIONING — separate global context from per-shot details (HoloCine pattern):

Prompt Structure (strict order):
1. GLOBAL: scene environment, time of day, weather, overall mood
2. CHARACTER DEFINITIONS: [char1]: full appearance anchor, [char2]: full appearance anchor
3. PER-SHOT: each panel references characters by tag [char1], never repeats full appearance

Why This Matters:
- Repeating character appearance in every panel prompt causes micro-variations → inconsistency
- Tag references ([char1]) force the model to maintain one canonical appearance
- Reduces total token count by ~40%, leaving more capacity for scene details

Per-Shot Format:
  Panel N: [shot type], [char1] does X while [char2] does Y, [lighting], [composition]

Anti-Patterns:
- DO NOT repeat "a young woman with long black hair wearing a red coat" in every panel
- DO NOT mix character definitions into per-shot descriptions
- DO NOT omit the global section — it anchors the entire sequence

## Example

GLOBAL: dim cyberpunk bar, night, rain outside, smoky neon mood.
CHARACTERS: [char1]: 30s woman, silver bob, scar on left cheek, black trench. [char2]: old bartender, bald, grey beard, maroon vest.
- Panel 1: wide, [char1] enters, [char2] polishes a glass, teal rim light.
- Panel 2: close-up, [char1] slides a coin, [char2] raises an eyebrow.
→ Appearance is never re-described per panel — only [char1] / [char2] tags carry it, saving tokens and locking one look.