image-gen

Backend-agnostic image generation workflow: single images, series with anchor-chain consistency, and batch pipelines. Two backends: Gemini (API) and Nano Banana (local scripts with post-processing).

Reference Loading Table

Phase 1: Detect Mode and Load References

Run the backend detection script — it reads environment variables and outputs a single word:

python3 skills/content/image-gen/scripts/detect-backend.py

Output values:

• gemini — GEMINI_API_KEY or GOOGLE_API_KEY is set
• ask — no key found; ask the user which backend to use

Load references based on output:

1. Load references/series-consistency.md (always — applies to every generation).
2. Load references/backend-selection.md (always — needed to pick mode and script).
3. Load references/backends/gemini.md when output is gemini.
4. Ask the user to set GEMINI_API_KEY or confirm they want to use local scripts when output is ask.

Gate: references loaded, backend confirmed before Phase 2.

Phase 2: Write Prompt File

Write the complete prompt to disk before any API call. Prompt files serve as the generation record and the anchor-chain input for series — writing them first means the full intent is on disk before any quota is spent.

File naming:

• Single image: prompts/YYYY-MM-DD-{slug}.md
• Series: prompts/{series-name}-01.md, prompts/{series-name}-02.md, ...

Prompt file format:

---
model: gemini-3-pro-image-preview
aspect-ratio: 1:1
flags: []
---

Full prompt text here. Be explicit about subject, style, background, and constraints.

Create the prompts/ directory if absent:

mkdir -p prompts

For a series, write all prompt files before calling any generation script. See references/series-consistency.md for the anchor-chain algorithm and why this ordering prevents drift.

Gate: all prompt files written and reviewed before Phase 3.

Phase 3: Select Mode and Script

Use references/backend-selection.md to map the request to the correct script and subcommand.

Gate: script and subcommand identified before Phase 4.

Phase 4: Generate

Call the selected script with absolute paths for output files — relative paths break when scripts run from different working directories.

For series generation, follow the anchor-chain sequence from references/series-consistency.md:

1. Generate image 1 with no reference.
2. Use output of image 1 as --reference for image 2.
3. Continue: each image references the previous output.

Show the full script output — the user needs status messages, warnings, and partial failure information.

Gate: script exits 0 before Phase 5.

Phase 5: Verify and Report

Visual inspection is mandatory. Read the generated image file to verify:

• Subject matches the prompt
• No unwanted watermarks, logos, or artifacts
• Aspect ratio and framing are correct
• No excessive padding or dark borders that need cropping

If visual inspection fails: regenerate with an adjusted prompt. Report the issue clearly before retrying.

Report to the user:

• Output file path (absolute)
• Image dimensions
• Model used
• Post-processing applied (if any)
• Visual verification result

Report only what was requested. The user did not ask for style suggestions or additional generations.

Error Handling