菜单
Generate illustrations in various styles — flat vector, watercolor, line art, 3D clay render, pixel art, and isometric. Auto-detects style from context. Triggers on: illustration, illustrate, flat design, watercolor, line art, pixel art, 3D illustration, isometric, sketch, cartoon, vector art, hand drawn.
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
Stages the changes produced in this session, excludes junk and secrets, and generates a conventional commit. Triggered when the user runs /devx-git:ci or asks to commit.
Updates the current branch onto an origin branch via rebase, with a recoverable backup, safe conflict resolution, and recovery steps. Use when the user runs /rebase or asks to rebase/update their branch onto main (or another origin branch).
Audits an agent, tool loop, or inference request across six dimensions (prompt engineering, tools definition, context engineering, cache usage, evaluation, guardrails) and produces a scored gap analysis with severity ratings and concrete remediations. Framework-agnostic; recognizes patterns from Anthropic SDK, OpenAI SDK, Vercel AI SDK, and OpenAI Agents SDK. Triggers on: audit my agent, review tool loop, audit inference request, agentic engineering review, audit prompt, review agent code, check prompt caching, evaluate guardrails.
Rewrites a given prompt or agent definition by applying Anthropic prompt-engineering best practices — clarity and directness, context and motivation, XML structure, role assignment, long-context layout, output format control, communication-style calibration, and parallel-tool-call affordance. Produces a before/after with per-change rationale. Triggers on: optimize this prompt, improve my prompt, rewrite prompt, prompt optimization, refine system prompt, tune agent prompt, polish prompt for production.
Analyzes project architecture and generates ASCII state machine and sequence diagrams. Use for explaining a project, analyzing architecture, tracing an entrypoint, showing a state machine, or generating a sequence diagram.
Audits a project's agentic harness (CLAUDE.md, AGENTS.md, .claude/rules, .claude/settings.json) and refactors it into a layered, path-scoped configuration that loads only what is relevant. Rewrites CLAUDE.md and AGENTS.md as a thin critical-rules layer, segments domain rules with `paths:` frontmatter, and adds deny-first permission guardrails. Use for: fixing the harness, auditing claude.md, refactoring agentic configuration, improving agent rules, segmenting rules by path, reviewing settings.json, harness review, or when claude.md is too long.
| name | banana-illustration |
| description | Generate illustrations in various styles — flat vector, watercolor, line art, 3D clay render, pixel art, and isometric. Auto-detects style from context. Triggers on: illustration, illustrate, flat design, watercolor, line art, pixel art, 3D illustration, isometric, sketch, cartoon, vector art, hand drawn. |
| allowed-tools | Bash(uv:*), Bash(ls:*), Read |
Input: $ARGUMENTS (optional — subject and style description)
If $ARGUMENTS is provided, use it as the illustration brief and skip to prompt building. If empty, ask about subject and style.
uv installedGEMINI_API_KEY environment variable set (get one at https://aistudio.google.com/apikey)[Subject/Character] + [Narrative Action] + [Illustrated Environment] + [Balanced Composition] + [Explicit Art Style]
| User Keywords | Resolved Style |
|---|---|
| "flat" / "vector" | Flat vector with solid fills, clean edges |
| "watercolor" | Wet-on-wet watercolor with paper texture, visible brushstrokes |
| "3D" / "clay" | Pixar-style clay render, soft shadows |
| "pixel" | Pixel art with visible grid, limited palette |
| "isometric" | Isometric projection, no perspective vanishing points |
| "line art" / "ink" | Consistent weight ink lines, minimal fills |
| Ambiguous | Default to flat vector |
| Domain | Aspect Ratio | Resolution |
|---|---|---|
| Spot / Icon | 1:1 | 1K |
| Scene | 4:3 | 1K |
uv run "${CLAUDE_PLUGIN_ROOT}/scripts/banana.py" \
--prompt "<detailed prompt following the formula>" \
--output "<YYYY-MM-DD-HH-MM-SS-descriptive-name>.png" \
--resolution 1K \
--aspect-ratio <ratio from Domain Defaults> \
--model gemini-3-pro-image-preview
When the user provides an existing image to restyle or modify:
uv run "${CLAUDE_PLUGIN_ROOT}/scripts/banana.py" \
--prompt "<edit instruction: what to change AND what to keep>" \
--output "<YYYY-MM-DD-HH-MM-SS-descriptive-name>.png" \
--input-image "/path/to/source.png" \
--aspect-ratio <ratio>
For style transfers, describe the target style and what to preserve:
"Recreate this photograph as a flat vector illustration with a limited palette of teal, coral, and cream. Keep the composition, subject pose, and environment layout exactly the same."
Up to 14 input images can be passed (repeat --input-image for each).
See references/prompts.md for ready-to-use example prompts covering common illustration styles and scenarios.