Image Poster Skill

Produce one finished image asset per turn unless the user asks for variations. Image generation rewards a tight, structured prompt — your job is to assemble that prompt from the user's brief, then dispatch.

Resource map

image-poster/
├── SKILL.md         ← you're reading this
└── example.html     ← what the resulting card looks like in Examples

Workflow

Step 0 — Read the project metadata

The active project carries imageModel, imageAspect, and (optional) imageStyle notes. Use them as the upstream model + canvas + style anchor; only ask the user to fill them in if they're marked (unknown — ask).

Step 1 — Compose the prompt

Plan in this exact order before calling any tool:

1. Subject + composition — what is in the frame, where, at what scale; eye-line and crop.
2. Lighting + mood — natural / studio / moody; warm / cool; key plus rim plus fill; time of day if outdoor.
3. Palette + textures — hex anchors when the user gave a brand palette; otherwise a 3-word mood tag (e.g. "muted ochre + ink").
4. Camera / lens — only if the user wants photographic realism ("85mm portrait, shallow DOF") or a specific film stock.
5. What to avoid — common AI-slop patterns ("no extra fingers, no warped text, no logo placeholders").

Step 2 — Dispatch via the media contract

Use the unified dispatcher — do not call upstream provider APIs by hand. Run from your shell tool:

node "$OD_BIN" media generate \
  --project "$OD_PROJECT_ID" \
  --surface image \
  --model "<imageModel from metadata>" \
  --aspect "<imageAspect from metadata>" \
  --output "<short-descriptive-name>.png" \
  --prompt "<the full assembled prompt from Step 1>"

The command prints one line of JSON: {"file": {"name": "...", ...}}. The daemon writes the bytes into the project folder; the FileViewer picks it up automatically.

Step 3 — Hand off

Reply with a one-paragraph summary of the prompt you used and the filename returned by the dispatcher (e.g. I generated hero-poster.png with gpt-image-2 at 1:1.). Do not emit an <artifact> tag.

Hard rules

• One image per turn unless asked for variations.
• Honor imageAspect exactly — the upstream cost is the same; matching the aspect avoids a re-render.
• No filler typography in the image itself unless the user asked for in-frame text. Real copy beats lorem.
• Save every render — never describe an image without producing the file. The user expects something to open in the file viewer.