AI Image Gen

AI Image Generation Skill

Single entry point for all image-generation tasks. Classifies the request, routes to the right path, and protects the user from paid API calls they didn't approve.

Hard rules

1. codex $imagegen is the default path. It costs nothing extra (uses the user's ChatGPT subscription login).
2. Never call a paid image generation path (Gemini MCP, OpenAI Images API, AI Studio MCP, etc.) without explicit user consent for that specific call. Previous consent does not carry over.
3. Never put file-management instructions inside the codex prompt (no "save as", "copy to", "confirm file exists"). This causes agent-loop stalls. The skill handles file movement after codex returns.
4. Never overwrite an existing asset unless the user explicitly asked for replacement. Use versioned filenames (-v2, -v3) instead.

Classification (do this first)

Read the user's request and classify into one of five categories:

Only A and D run without asking. B, C, E always confirm first.

Path: codex $imagegen (default for A and D)

Prompt rules

• Prefix the prompt with $imagegen (shell-escape as \$imagegen when calling from bash to avoid variable expansion).
• Include only the visual description. Do not include save paths, filenames, or confirmation steps.
• If the user's original request contains heavy file-management instructions, strip them from the codex prompt and handle them yourself.
• Keep the visual prompt reasonable length. If the user's prompt is very long and packed with negative conditions, it can stall the codex agent loop — consider light trimming or warn the user.

Invocation

codex exec --full-auto --skip-git-repo-check "\$imagegen <pure visual prompt>"

• Wall-clock timeout: 600 seconds (10 minutes).
• If 60 seconds pass with codex still running, emit a brief "仍在產圖中…" progress note to the user so they know it's not hung.
• For edit mode (category D), pass the source image via -i <path> to codex exec, and phrase the prompt as an edit instruction.

Locating the output

After codex finishes:

1. Parse session id: <UUID> from codex stdout.
2. Check ~/.codex/generated_images/<session-id>/ig_*.png.
3. If parse fails, fall back to: most recently modified subdirectory under ~/.codex/generated_images/.
4. If still no ig_*.png exists → treat as failure, go to the failure-fallback section.

Copy to target

• If user specified a destination filename/path → use it.
• If not specified → default to the current working directory with name <semantic-slug>-<YYYYMMDD>.png (extract a short slug from the prompt, e.g. superhero-20260417.png).
• If the target filename already exists → append -v2, -v3, … until unique.
• Use cp, not mv — leave the original in ~/.codex/generated_images/ alone.

Path: ask user first (for B, C, E)

Present the choice concisely, then wait:

• B (logo/icon): "這看起來是 logo/icon，建議走 SVG/HTML 直接寫（可編輯、體積小、無成本）；也可以用 codex $imagegen 產 raster。要走哪一條？"
• C (diagram): "這像是流程圖/架構圖，建議用 Mermaid 或 SVG 結構化（可編輯、可版控）；也可以用 codex $imagegen 產 raster。要走哪一條？"
• E (unclear): Describe your best guess of the categories and ask the user to pick.

Proceed only after user picks. If they pick codex raster, follow the codex path above.

Failure fallback (paid path — requires explicit consent)

Codex path is considered failed when:

• Wall-clock exceeds 600 seconds → kill codex, proceed to fallback.
• Codex finishes but no ig_*.png was produced in the session directory.
• Codex returns a non-zero exit with an obvious error (auth, quota, sandbox).

On failure:

1. Report the failure mode to the user (what happened, briefly — timeout vs no output vs error).
2. Ask explicitly, in one turn: "Codex 失敗（原因 X）。要改用 Gemini MCP 嗎？會扣 GEMINI_API_KEY（約 $0.03–0.06/張）。要繼續就回 yes。"
3. Wait for explicit "yes" (or equivalent clear consent). Do not infer consent from silence, context, or prior preferences.
4. On consent: call mcp__gemini__gemini-generate-image. Copy MCP output to target destination using the same naming rules as the codex path.
5. If the user declines or stays silent → stop, leave the task incomplete, do not try other paid paths.

Other paid paths (mcp__aistudio__generate_content, OpenAI Images API direct) follow the same consent rule — never call without a per-call explicit yes.

Error modes to watch for

• codex agent loop stall: seen with long complex prompts that also include file-management instructions. Mitigation is in the "Prompt rules" section above — keep the codex prompt pure-visual.
• stdout pipe buffering: do not pipe codex output through | tail -N. It can buffer and hide progress. Let stdout stream directly.
• stale superhero.png etc.: if a previous codex run partially failed, old files may be missing or in unexpected state. Never assume target dir is clean — check before writing.
• $ expansion: in bash, $imagegen will expand to empty string. Always escape as \$imagegen in the shell command.

Example (category A, default path)

User: "幫我產一張賽博朋克風的街道夜景"

1. Classify → A (scene/style keywords).
2. Build codex prompt: $imagegen A cyberpunk street scene at night, neon lights, rainy asphalt, reflections, dense urban atmosphere, cinematic wide angle
3. Run codex exec --full-auto --skip-git-repo-check "\$imagegen …".
4. Parse session id, locate ig_*.png, cp to ./cyberpunk-street-20260417.png.
5. Report path + size to user.

Example (category B, ask first)

User: "幫我做一個公司 logo，藍色簡潔風"

1. Classify → B.
2. Ask: "logo 建議走 SVG/HTML 直接寫，可編輯、體積小、無成本；也可以用 codex $imagegen 出 raster。要哪條？"
3. Wait.
4. If user picks codex → follow default path. If SVG → write SVG directly, don't use codex.