Menu
Generate a new image from a text prompt using Gemini or OpenAI.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Run quality checks on a task's implementation before it closes to done. Triggers: 'is this ready?', 'run the review gate', 'check my work', 'I think this is done'. Reviews code + spec adherence, runs tests/build. Design review uses taskmaster:spec-review.
Log a lightweight idea into .taskmaster/ideas/. Triggers: '/add-idea', 'save this as an idea', 'remember this idea', 'log this idea', or something worth keeping, not yet a task. Only way - not backlog_idea_create directly.
Log/update/close project bugs in .taskmaster/bugs/. Triggers: 'log a bug', 'this is a bug', 'track this defect', 'I found a bug', 'shelve this for later', 'list open bugs', 'promote B-XX', 'close B-XX'. Also for ambiguous 'issue' phrasing lacking recurring/systemic/outstanding evidence - falls back here. Only correct way to transition a project Bug.
Scan the codebase for TODO/FIXME/HACK/XXX and cross-reference the backlog. Triggers: 'check todos', 'are my todos tracked', 'scan for todos', 'todo audit', 'what's untracked'.
Write/resolve/drop project decisions in .taskmaster/decisions/. Invoke when about to write an inline option menu - route here, not Options: in chat. Also 'choose between', 'pick an option', 'decide on', 'open question', 'resolve DEC-X', 'drop DEC-X'. Only correct way - do not call backlog_decision_create directly.
Close out a work session. Triggers: 'end session', 'I'm done for today', 'let's wrap up', 'mark this task done', 'save progress'. ONLY correct way to mark tasks done/in-review with a session record.
| name | generate |
| description | Generate a new image from a text prompt using Gemini or OpenAI. |
Create a new image from a text description.
Craft the prompt — Transform the user's request into a detailed image generation prompt. Be specific about style, composition, colors, lighting, and subject matter. The more detail, the better the result.
Choose backend and parameters — Based on the use case:
--backend openai --transparent (native alpha — script auto-uses gpt-image-1.5 because gpt-image-2 does not support transparency) or --backend gemini --transparent (two-pass extraction, slower)--model gemini-3-pro-image-preview for final quality, --backend openai --quality high for best OpenAI quality (defaults to gpt-image-2)Choose output path — Name the file descriptively based on what it depicts. Place in ./assets/generated/ by default, or in a project-appropriate location (e.g., ./public/images/, ./src/assets/).
Run the script:
# Standard generation (Gemini)
node "${SKILL_DIR}/../../src/generate.mjs" \
--prompt "<detailed prompt>" \
--output "<output-path>.png" \
--aspect "<ratio>" \
--size "<size>"
# Transparent background (OpenAI — recommended)
# Script auto-uses gpt-image-1.5 because gpt-image-2 does not support transparency.
node "${SKILL_DIR}/../../src/generate.mjs" \
--backend openai \
--transparent \
--prompt "<detailed prompt>" \
--output "<output-path>.png"
# Transparent background (Gemini two-pass — slower, 2 API calls)
node "${SKILL_DIR}/../../src/generate.mjs" \
--transparent \
--prompt "<detailed prompt>" \
--output "<output-path>.png"
View the result — Use the Read tool to view the generated image file.
Report back — Show the user the image path and describe what was generated. Ask if they want adjustments.
| Scenario | Recommendation |
|---|---|
| Game sprites, icons, UI elements | --backend openai --transparent (auto-uses gpt-image-1.5) |
| Transparent + specific Gemini style | --backend gemini --transparent (two-pass) |
| Backgrounds, photos, art | No --transparent needed (uses gpt-image-2) |
Important — model split for OpenAI:
gpt-image-2 (default for non-transparent) is the newer, state-of-the-art model with flexible sizes and faster generation. It does NOT support transparent backgrounds — this is a regression vs. gpt-image-1.5.gpt-image-1.5 is retained specifically for transparency. The script automatically falls back to it when --transparent is set, so you usually don't need to pass --model yourself. If you do pass --model gpt-image-2 --transparent together, the script errors out instead of silently failing at the API.Gemini transparency: Gemini cannot produce true alpha channels. When --transparent is used with Gemini, the script generates the image twice (on white and black backgrounds) and computes alpha mathematically. This uses 2 API calls and may have artifacts on complex edges.
GEMINI_API_KEY is not set, tell the user to set it: export GEMINI_API_KEY=your-keyOPENAI_API_KEY is not set for OpenAI backend: export OPENAI_API_KEY=your-keynpm install --prefix "${SKILL_DIR}/../../src"