菜单
Use this skill to create, generate, draw, or edit images/photos into saved PNG files with OMLX image models. Covers text-to-image and image-to-image edits. Do not use for text-only image explanations, screenshot summaries, or non-generative image processing scripts.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Automate browser interactions, test web pages and work with Playwright tests.
Coordinate durable coding-job handoffs between Microsoft Scout and GitHub Copilot through a shared SQLite queue. Use when asked to enqueue work for Copilot, claim or pick up queued work, identify the repository for a claimed job, mark or safely unblock blocked work, release a claimed job for retry, record a pushed branch and draft pull request, or discover blocked and finished jobs. Do not use for ordinary coding work that was not queued or for generic process scheduling.
Use for performing Azure DevOps work — create, inspect, review, comment on, vote on, or manage pull requests; find, query, create, update, and link Azure Boards work items (including building WIQL queries); parse and route dev.azure.com or *.visualstudio.com URLs; and upload PNG/image/file attachments to pull requests. Triggers on Azure DevOps / ADO / Azure Boards links or action requests such as "make an ADO PR", "review this Azure DevOps PR", or "find work items assigned to me". Not for general conceptual explanations that do not act on a specific Azure DevOps resource, and not for GitHub, Jira, or other non-Azure-DevOps tools.
Use this skill to turn a video or audio recording (a talk, meeting, or demo) into written content: documentation, a blog post, tutorial, changelog, or notes. Handles transcription, takeaway synthesis, screenshot/frame selection, and grounded drafting from local OpenAI-compatible media models. Not for real-time transcription, generic video editing, or non-generative media processing.
Use this skill to create, revise, package, evaluate, or optimize GitHub Copilot skills and reusable SKILL.md-based agent instructions. Use it whenever the user wants a new Copilot skill, edits to an existing skill, eval/test prompts, skill benchmarking, skill packaging, trigger-description optimization, or reusable agent instructions.
Create or improve AGENTS.md and project agent-skill guidance for a repository. Use when the user asks to initialize, bootstrap, or set up agent instructions, repo guidelines, AGENTS.md, or reusable skills. Discovers repo structure, commands, existing instruction files, and .agents/skills, then writes concise, high-signal guidance.
| name | image-gen |
| description | Use this skill to create, generate, draw, or edit images/photos into saved PNG files with OMLX image models. Covers text-to-image and image-to-image edits. Do not use for text-only image explanations, screenshot summaries, or non-generative image processing scripts. |
| allowed-tools | Bash(uv run ./scripts/generate.py:*) Bash(uv run ./scripts/edit.py:*) |
$OMLX_BASE_URL must be set.uv must be available to run the bundled Python helpers..png, .jpg, .jpeg, or .webp file.--input, --inputs, --mask, and --output.$OMLX_BASE_URL/v1/models/status unless the user specified one. The response puts models under .models (not .data). Prefer models with engine_type or model_type of image and a loaded status. A working discovery query:curl -s "$OMLX_BASE_URL/v1/models/status" | jq '.models[] | select(.engine_type == "image" or .model_type == "image") | {id, loaded, capabilities}'
Prefer loaded models whose capabilities or tasks include generation or edit.
3. Use uv run ./scripts/generate.py or uv run ./scripts/edit.py. Do not hard-code install locations, copy scripts elsewhere, or run non-bundled wrappers. The scripts refuse outputs inside the skill directory.
4. For structural edits that transform a scene while preserving the subject, prefer FLUX.2/Klein edit models over ERNIE when both are available. Prompt explicitly for preserved subject details (pose, clothing, crop, camera angle, privacy blur) and the replacement scene. If the first result changes too much or too little, retry with --image_strength values such as 0.2, 0.4, or 0.6, and use --steps/--guidance for quality and prompt-adherence sweeps.
5. Report the saved file path and model used.
Generate:
uv run ./scripts/generate.py \
--prompt "a watercolor fox in a moonlit forest" \
--model "<image-model>" \
--size "1024x1024" \
--output "/absolute/path/to/user/workspace/result.png"
Edit:
uv run ./scripts/edit.py \
--input "/absolute/path/to/user/workspace/source.png" \
--prompt "add sunglasses" \
--model "<image-edit-model>" \
--output "/absolute/path/to/user/workspace/edited.png"
Read references/generation.md or references/editing.md only when you need optional parameters such as --n, --quality, --style, --mask, --steps, --guidance, --image_strength, multi-input edits, or URL responses.