Files

T

pftom 976a6eadf2 feat(media): add image/video/audio project kinds via od media generate

Introduce non-web media surfaces (image, video, audio) as first-class
project kinds. The unifying contract is "skill workflow + project
metadata tell the agent WHAT to make; one shell command — od media
generate — is HOW bytes are produced", so any code-agent CLI with
shell access can drive it without bespoke tools.

- Frontend: New Project panel gains Image/Video/Audio tabs with model
  picker, aspect/length/duration controls, and audio kind/voice
  selection. Examples and Design Systems tabs gain layered sections.
  FileViewer renders the generated image/video/audio files.
- Shared registry: src/media/models.ts is the single source of truth
  for image/video/audio model IDs, aspects, and defaults — consumed
  by the picker AND the daemon dispatcher.
- Prompts: media-contract.ts is pinned LAST in the system prompt for
  media surfaces so its hard rules (call od media generate, don't
  emit binary in <artifact>, allowed model IDs) win over softer
  earlier wording.
- Daemon: new media.js dispatcher + media-models.js JSON view of the
  registry; cli.js gets the `od media generate` subcommand wired up
  via server.js / projects.js so the daemon writes files back into
  the project dir.
- Skills: audio-jingle, image-poster, video-shortform seed examples
  for the three surfaces.

Made-with: Cursor

2026-04-28 22:41:14 +08:00

3.4 KiB

Raw Blame History

name, description, triggers, od

name

description

triggers

image-poster

Single-image generation skill for posters, key art, and editorial illustrations. Defaults to gpt-image-2 but is provider-agnostic — the same workflow drives Flux, Imagen, or Midjourney via the active upstream tooling. Output is one or more PNG/JPEG files saved to the project folder.

poster

key art

illustration

image

cover art

海报

插画

mode

surface

scenario

preview

design_system

example_prompt

image

design

type	entry
html	example.html

requires
false

Editorial poster for an indie film festival — one bold abstract silhouette over a warm, slightly grainy paper background; hand-set sans serif title at the top, festival dates and venue at the bottom in monospace. Muted ochre + ink palette.

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:

Subject + composition — what is in the frame, where, at what scale; eye-line and crop.
Lighting + mood — natural / studio / moody; warm / cool; key plus rim plus fill; time of day if outdoor.
Palette + textures — hex anchors when the user gave a brand palette; otherwise a 3-word mood tag (e.g. "muted ochre + ink").
Camera / lens — only if the user wants photographic realism ("85mm portrait, shallow DOF") or a specific film stock.
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.

3.4 KiB Raw Blame History