New Image Workflow

Scaffold a structured, multi-stage image workflow inside the user's registered image workspace. Unlike new-project (a flat source/working/exports layout), a workflow expects iteration: generation, human review, revision, and export are explicit stages, each with their own folder and manifest entry.

Cloud AI (Fal nano-banana via the fal REST queue API) is wired in as an optional driver for the generation and revision stages.

When to use this vs. new-project

• new-project — single batch of images, photo-editing style, no iteration loop.
• new-workflow — generative or hybrid work with multiple touchpoints: brief → references → generate → review → revise → export. Use when the user expects to come back to the project across several sessions and wants a stable structure to resume into.

Procedure

1. Resolve the workspace

CONFIG="${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/image-production/workspace.json"
test -f "$CONFIG" || { echo "No workspace registered — run workspace-setup first."; exit 1; }
WS=$(jq -r .path "$CONFIG")

2. Gather workflow metadata

Ask (or infer if obvious):

• Name — required. Slugify (lowercase, hyphens).
• Brief — one or two sentences describing the deliverable.
• Cloud-AI driver — default fal-nano-banana (via the fal REST queue API, see skill ai-graphics). none if the user is bringing their own images.
• Expected rounds — how many revision cycles to scaffold (default 3).
• Final aspect ratio / resolution — for export-stage placeholders.

3. Create the layout

PROJ="$WS/<slug>"
test -e "$PROJ" && { echo "Already exists: $PROJ"; exit 1; }
mkdir -p "$PROJ"/{01-brief,02-references,03-generation,04-review,05-revision,06-export,_manifest}
mkdir -p "$PROJ"/05-revision/round-{1..N}   # N = expected rounds

4. Write the manifest

$PROJ/_manifest/workflow.json:

{
  "slug": "<slug>",
  "created": "<ISO date>",
  "brief": "<one-line summary>",
  "driver": "fal-nano-banana | none",
  "stages": [
    {"id": "01-brief",       "status": "pending", "type": "human"},
    {"id": "02-references",  "status": "pending", "type": "human"},
    {"id": "03-generation",  "status": "pending", "type": "ai"},
    {"id": "04-review",      "status": "pending", "type": "human"},
    {"id": "05-revision",    "status": "pending", "type": "ai+human", "rounds": N, "current_round": 0},
    {"id": "06-export",      "status": "pending", "type": "ai"}
  ],
  "final": {"aspect_ratio": "<...>", "resolution": "1K", "format": "png"}
}

5. Seed templates

• 01-brief/brief.md — markdown template with sections: Goal, Audience, Deliverables, Constraints, Do, Don't, Reference Links.
• _manifest/prompts.md — append-only log; one heading per generation call with the prompt, model params, and resulting filenames.
• README.md at project root — slug, brief, driver, ISO date, stage checklist (mirrors workflow.json), and a "Resume here" pointer to the first non-done stage.

6. (Optional) git init

If the workspace itself is a git repo (check git -C "$WS" rev-parse --git-dir), the new workflow is automatically tracked — no separate init. Otherwise, ask whether to git init the project folder.

7. Confirm and hand off

Report:

• Project path
• Current stage (01-brief)
• Next action (open 01-brief/brief.md and fill in)

Driving the workflow

Subsequent sessions resume by reading _manifest/workflow.json and continuing from the first stage with status != "done".

Cloud AI generation (stage 03 + 05)

When the active stage is ai-typed and the driver is fal-nano-banana:

1. Read the brief + references; compose a dense one-paragraph prompt (see ai-graphics skill for prompt style).
2. Submit to https://queue.fal.run/fal-ai/nano-banana-2 with prompt, image_size (enum: square, landscape_4_3, landscape_16_9, portrait_3_4, portrait_9_16; no resolution parameter) and output_format, then poll and fetch as in skill ai-graphics ("Run it"). For revision rounds that build on a prior image, use fal-ai/nano-banana-2/edit with image_urls (URLs or data: URIs, so a local file can be base64-encoded inline), falling back to fal-ai/flux-2/edit if rejected.
3. Download the returned image URL into the stage folder:

   curl -sSL -o "$PROJ/03-generation/$(date +%s)-<slug>.png" "<fal URL>"

4. Append a log entry to _manifest/prompts.md: stage, round, prompt, params, output filename.
5. Mark the stage awaiting-review in workflow.json and surface the candidates to the user.

Human touchpoints (stage 01, 02, 04, 05-review)

At each human stage, stop and prompt the user. Do not auto-advance. Show:

• What was just produced (file list, thumbnails if available)
• The decision needed (which candidates to keep, what to change, etc.)
• Where to record the decision (a notes.md in the stage folder)

After the user records the decision, mark the stage done and advance.

Export (stage 06)

When all revision rounds are done:

1. Take the selected final images from the last 05-revision/round-N/ folder.
2. Apply: resize to target resolution, convert to target format, scrub metadata (use optimize-jpeg / convert-to-webp / scrub-metadata skills as appropriate).
3. Write to 06-export/ with stable filenames (<slug>-<variant>.<ext>).
4. Mark stage done.

Notes

• Don't overwrite existing files in 03-generation/ or 05-revision/round-*/ — always timestamp-prefix.
• The manifest is the source of truth; the README is a human-readable mirror, regenerate it from the manifest when state changes.
• If the user changes the driver mid-flow (e.g., switches from cloud-AI to bringing their own), update workflow.json and note it in prompts.md.