Skill: Two-Layer Scene Direction with Identity Anchors
What
Stage 1 creates identity anchors (neutral grey, no backgrounds). Scene direction uses two layers: the original plan (written upfront) and evolved prompts (adapted after seeing each panel's actual output). Both are stored.
Why
Writing all prompts upfront and generating blindly fails — you can't predict what the model will produce. But having NO plan means you're improvising without a compass.
The two-layer approach gives you both:
- Original scenes = your narrative compass (story arc, beats, intent)
- Evolved prompts = your actual path (adapted to what the model produced)
The gap between them is learning data.
How
Stage 1: Identity Anchors (neutral grey only)
Generate character/object reference sheets on NEUTRAL GREY BACKGROUNDS ONLY:
- Character turnaround grids (front, back, side, 3/4)
- Expression sheets (key emotions)
- Motion sheets (key poses) — if needed
- Room reference views (these DO show the environment)
CRITICAL: No environments/backgrounds on character sheets. Ever.
Pre-Production: Write the Original Scene Plan
Write all scene descriptions upfront in scene-plan.md. This is your intended direction — the narrative beats, camera angles, emotional arc. This file is your compass and is never modified during generation.
Panel Generation: Sequential + Adaptive
1. Read original scene 1 → write prompt → generate panel 1
2. SEE panel 1 — what did the model actually produce?
3. Read original scene 2, ADAPT it based on panel 1's actual output → generate panel 2
4. SEE panel 2 — what changed? What carried forward?
5. Read original scene 3, ADAPT based on panel 2 → generate panel 3
...repeat for every panel
Key behaviors:
- Always reference the original scene plan — it tells you where the story SHOULD go
- Adapt each prompt to the reality of the previous panel's output
- If a character ended up in an unexpected position → write the next scene from where they actually are
- If the model gave you something better than planned → follow it
- If something went wrong → regenerate before continuing (chain is broken)
Store Both Layers
scene-plan.md — original intent, never touched during generationprompts.json — evolved prompts, what was actually sent to the model
The divergence between these two documents is diagnostic data. It tells you:
- Where the model's interpretation differed from yours
- How you adapted (and whether the adaptation was good)
- What kinds of scenes need tighter direction vs. which ones the model handles well
File Structure
projects/{id}/
scene-plan.md ← original plan (compass, never modified)
prompts.json ← evolved prompts (adapted during generation)
objects/
daughter/
identity-anchor.png (neutral grey)
emotion-anchor.png (neutral grey)
hospital-room/
view-door.png (environment views — these show the room)
No scene-refs/ folder. Scene-ref batches are eliminated.
Common Mistakes
- No original plan — improvising without a compass. Write the full scene plan first, THEN adapt during generation.
- Modifying scene-plan.md during generation — the original is your reference. Write evolved prompts to prompts.json instead.
- Not looking at output before writing next prompt — the whole point is reactive direction. See it, then write the next one.
- Fighting the model's output — if the character ended up facing left, write the next scene from that reality. Don't force a 180 unless the story requires it.
- Backgrounds on reference sheets — neutral grey only. No environments on character/motion/expression refs.
- Pre-generating scene-ref batches — blind prediction. Eliminated.
Lessons from hospital-entry-test
- Pre-planned 14 prompts + 4 scene-ref batches upfront
- Scene-refs had backgrounds/environments (wrong — neutral grey only)
- No feedback loop between generation and scene writing
- Cost: $2.53 for a result that didn't hold consistency
- Learning: need two layers (plan + evolved) and sequential generation with adaptation