trae-agent-enhancements

Look up AI model specs from official sources and output a configuration table for Trae. Use when the user asks to add/configure a model or look up context window, including Chinese requests such as 添加模型, 模型怎么填, 上下文窗口是多少, 模型参数, or 帮我查模型信息. Also covers requests mentioning specific model names like DeepSeek, Claude, GPT, Gemini, or Qwen.

Manage and query persistent cross-session memory via MCP Knowledge Graph. Step 0 executor: query before ANY task for project/pattern/user context. Use when the task involves remembering, retrieving, or storing project knowledge, or when you need historical context, including Chinese requests such as "你还记得吗""之前讨论过""校准记忆""同步记忆". Does NOT replace Trae Core Memory — complements it.

Operate GitHub repositories, pull requests, issues, and releases via the `gh` CLI. Use when the user asks to push, create PR, manage issues, edit repo metadata, view repo, or any GitHub workflow operation, including Chinese requests such as “推送”“提 PR”“创建 issue”“改仓库描述”“查看仓库信息”.

Configure Git for domestic platforms and manage branching strategies suitable for Chinese teams. Use when the user asks to configure remotes, set up SSH, or manage Git workflows for Gitee, Coding, Jihu GitLab, or CNB, including requests like "配置国内 Git", "设置 Gitee 远程仓库", "推荐适合小团队的 Git 流程".

Use when there are 2 or more independent tasks that can be delegated to available Task subagents in parallel, including Chinese requests such as “并行搜索”“并行审查”“并行浏览器验证”“并行跑子代理”. Do not use when tasks share file edits, depend on shared state, require sequential reasoning, or are faster with direct tools.

Use when you need a safe git worktree before starting isolated feature work, experiments, or plan execution, including Chinese requests such as “开隔离工作区”“不要污染当前分支”“先建 worktree 再做”. Do not use for read-only analysis, tiny edits that do not justify a separate workspace, or when the user explicitly wants to work in the current tree.

Uses Chrome DevTools MCP for accessibility debugging and auditing in Trae. Use when testing semantic HTML, ARIA labels, accessible names, focus states, keyboard navigation, tap targets, color contrast, Lighthouse a11y, or Chinese requests such as 可访问性检查, 无障碍审计, 键盘导航测试, 表单标签检查, 按钮没有名称, or 颜色对比度检查.

Use when the user asks to brainstorm, clarify requirements, compare approaches, design a feature, draft a spec, sketch a prototype, or think through architecture before implementation. 适用于需求不清、方案未定、需要先梳理思路再实现，以及功能设计、组件设计、页面结构、交互流程、数据流和模块边界讨论。先探索用户意图、约束和设计，再决定是否进入实现。

Guides LCP and Core Web Vitals debugging in Trae with Chrome DevTools MCP performance tools. Use for Largest Contentful Paint, slow page loads, hero image render delays, page load speed, CWV, Performance trace, or Chinese requests such as 优化 LCP, 页面加载慢, 首屏慢, 核心网页指标, 性能分析, 主内容渲染太慢, or 首屏图片加载慢.

Use Everything Search MCP in Trae for fast Windows local file and folder path discovery. Use when the user needs to find local files, folders, projects, PDFs, images, documents, archives, or recently modified items by name, extension, path fragment, date, or size, including Chinese requests such as 找文件, 搜索本机文件, 定位项目目录, 查找某个 PDF, 最近修改的文件, or 文件在哪. Prioritize this over default shell commands (dir/ls) or the Glob tool when searching outside the workspace, across the whole machine, or by metadata. If a workspace Glob search fails, fallback to this Skill. Do not use for semantic code search, web search, or project understanding (use the 'search' agent instead).

Use when implementation is done, validation has passed, and you need to decide how to wrap up the current branch, including Chinese requests such as “接下来合并还是提 PR”“这条分支怎么收尾”“保留还是丢弃当前工作”. Do not use while implementation or verification is still in progress, or when the user only wants a normal commit.

Execute git commits in Trae with conventional commit message analysis, intelligent staging, and message generation. Use when the user asks to commit changes, create a git commit, mentions /commit, or says Chinese requests such as 帮我提交, 生成提交信息, 提交当前修改, or 按规范提交. Supports type/scope detection, conventional commit messages, and logical file staging.

Captures learnings, errors, and corrections to enable continuous improvement. Use when a command fails unexpectedly, the user corrects a misunderstanding, an external tool breaks, or a better approach is discovered. 适用于“把这个报错记下来”、“记住我刚才说的规则”、“更新最佳实践”或“把这个教训沉淀下来”等需要自我进化的场景。

Create, modify, refactor, review, or iterate Trae Skills. Use only when the user is explicitly working on a Skill or SKILL.md, including Chinese requests such as 创建 skill, 修改 SKILL.md, 优化触发条件, 适配 Trae, 清理旧平台内容, or 减少歧义. Do not use for ordinary coding, docs writing, prompt writing, workflow discussion, or project refactoring unless the user asks to turn that work into a Skill or edit an existing Skill.

Use only when the user wants to review Trae Skill packages or SKILL.md files, including related scripts, resources, references, templates, config examples, trigger boundaries, Windows/Trae adaptation, Chinese-user fit, execution stability, script synchronization, and over-localization risks inside a Skill package. Typical Chinese requests include "审查这个 skill", "扫描 skills", "检查这个 SKILL.md", "评估这个 Trae Skill 是否稳定", or "检查 skill 的脚本同步适配". Do not use for ordinary project review, code review, repository readiness, git/GitHub push standards, release checks, general script checks, general Chinese writing polish, or non-Skill stability reviews unless the target is explicitly a Trae Skill package or SKILL.md.

Use when you need to turn a clear spec, approved design, or stable requirement into an executable multi-step implementation plan before coding, including Chinese requests such as “写实现计划”“拆任务”“先规划再写代码”. It is especially suitable for handing off a brainstorming result into file-level tasks, tests, and validation steps. Do not use when requirements are still unclear, the change is small enough to implement directly, or you are already executing an existing plan.

Use when the user needs to design, create, optimize, split, or name a reusable Trae agent configuration for a concrete task, including Chinese implicit requests such as "需要一个能做 X 的专门助手/agent/智能体", "现有 agent 职责太宽或能力不够", "这个功能是否应该独立成 agent", or "想把某个工作流固定成可复用 agent". Do not use for ordinary prompt writing, Trae Skill/SKILL.md creation or review, software architecture, tool selection, general agent discussion, or requests without intent to produce a reusable Trae agent configuration.

Apply professional Chinese copywriting and typography guidelines. Use when the user asks to format, polish, write, or translate technical documentation, plans, code comments, or reports in Chinese, including requests like "排版一下文档", "中文文档规范", or "调整中英文混排".

Use when you need to execute a written implementation plan in the current conversation with review checkpoints, including Chinese requests such as “按这个计划执行”“照计划做”“逐任务推进”. It is especially suitable after writing-plans or when the environment does not support implementation-capable subagents. Do not use when requirements are still unclear, no written plan exists, or subagent-driven-development is a better fit for mostly independent tasks.

Use when you already have code review feedback and need to validate, clarify, and apply it safely, including Chinese requests such as “这些 review 意见怎么处理”“收到审查反馈后怎么落地”“先别急着改，帮我判断哪些该修”. Do not use when no review feedback exists yet, or when the task is ordinary code review rather than responding to comments.

Use when you need an independent code review before commit, merge, or task handoff, including Chinese requests such as “帮我 review”“合并前审一下”“看看这次改动有没有问题”. Do not use for from-scratch implementation work with no reviewable diff yet, or when the user only wants a code explanation instead of a risk-focused review.

Use when debugging bugs, test failures, build breaks, performance regressions, or unclear behavior before proposing a fix, including Chinese requests such as “逐步排查”“为什么失败”“先定位原因”“不要先猜修复”. Do not use for straightforward implementation tasks, pure document edits, or issues that already have a confirmed root cause and only need mechanical changes.

Use when implementing behavior with tests first, including new features, bug fixes, behavior changes, and regression protection, and including Chinese requests such as “先写测试”“补回归”“按 TDD 修复”. Do not use for pure copy edits, generated outputs with little test value, or tasks where the user explicitly wants to skip testing and accepts the risk.

Run an agency-orchestrator style YAML workflow inside the current Trae conversation by simulating or coordinating multiple roles. Use when the user provides a `.yaml` or `.yml` workflow, asks for multi-role collaboration, says things like “用多个角色一起评审这个 PRD”“多角色协作”“先生成工作流再执行”“让产品经理、架构师、QA 一起看一下”, or wants roles such as product manager, architect, reviewer, or QA to work together. Supports layered execution, role prompting, variable interpolation, document-backed inputs, and parallel steps only when the current environment has a clear subagent mapping. Do not use for ordinary implementation plans, simple code changes, or tasks already covered by `writing-plans`, `executing-plans`, `subagent-driven-development`, or single-purpose review/debug skills.

Use this skill to discover available subagents and their capabilities by reading the system prompt's Task tool definition. Use when the user asks what agents are available, what a specific agent can do, how to find the right agent for a task, whether there is an agent for X, or how to understand agent capabilities, including Chinese requests such as “有哪些子代理可用”“这个 agent 能做什么”“怎么选择合适的 agent”. Do not use for Custom Agent discovery (encrypted database, no read access).

Use before claiming work is done, fixed, tested, ready to commit, or ready to merge, including Chinese requests such as “确认真的修好了”“提交前再验证一下”“不要没跑就说完成”. Re-run the proving checks and report the real status from exit codes and output. Do not use while work is still exploratory or when no completion claim is being made.

Visualize data in Trae by selecting the most suitable chart type, extracting parameters, and generating a chart image with the local JavaScript script. Use for Chinese or English requests such as 数据可视化, 生成图表, 画折线图, 柱状图, 饼图, 漏斗图, 词云, 地图, 组织架构图, or 把这些数据做成图.

Uses Chrome DevTools MCP in Trae for real browser automation, webpage debugging, Console/Network/DOM inspection, performance analysis, and UI flow verification. Use for Chinese or English requests such as 调用浏览器, 打开网页操作, 自动点击填写提交, 检查网页报错, 分析 Network 请求, 页面性能分析, or Chrome DevTools debugging. This skill does not apply to slim MCP mode.

Create, modify, review, or organize Trae IDE rules that live under .trae/rules, module-level .trae/rules folders, AGENTS.md, CLAUDE.md, or git message rule files. Use when the user asks to 创建 rule, 创建规则, 编写 Trae 规则, 适配 Trae rules, set alwaysApply/globs/description/scene, create commit message rules, split project rules by module, or convert team conventions into reusable Trae project rules.

Retrieve current documentation, API references, configuration guidance, migration notes, and code examples for developer technologies through Context7 in Trae. Use when the user asks about a library, framework, SDK, API, CLI tool, or cloud service, including Chinese requests such as 怎么配置, 如何使用, 报错原因, 版本迁移, API 参数, or 安装步骤. Prefer this over general web search for developer documentation.

Create distinctive, production-grade frontend interfaces in Trae with high design quality. Use when the user asks to build or polish web components, pages, dashboards, React/Vue/HTML/CSS UI, landing pages, posters, or Chinese requests such as 美化页面, 优化前端界面, 做一个网页, 设计组件, 调整样式, or 提升 UI 质感. Generates polished code and avoids generic AI aesthetics.

Process and manipulate images in Trae with Python orchestration and ImageMagick on Windows-friendly paths. Use when the user needs reusable image automation, batch conversion, resizing, cropping, padding, metadata inspection, or Chinese requests such as 批量处理图片, 批量转格式, 裁剪图片, 补白边, 生成缩略图, 压缩成 webp, or 批量改尺寸.

Diagnose JavaScript, browser, and Node.js memory leaks in Trae with Chrome DevTools MCP and memlab-oriented workflows. Use when the user reports high memory usage, OOM errors, heap snapshots, detached DOM nodes, growing memory, or Chinese requests such as 内存泄漏, 内存占用高, 页面越用越卡, 分析 heapsnapshot, or 排查 Node 内存问题.

Convert supported local documents to Markdown through the local MinerU API in Trae, including PDF, images, DOCX, PPTX, and XLSX, and save the output to a chosen Windows-accessible directory. Use when the user asks to 转 Markdown, 提取 PDF, 解析文档, 图片转 md, 保存为 markdown, or batch-convert supported files. Do not use for web pages, plain text rewriting, or summarization tasks without file conversion.

Define the repository-wide language policy for Trae Skills. Use when creating, refactoring, or reviewing Skills that need consistent rules for English machine-facing structure, retained Chinese trigger phrases, and agent-first language decisions, including Chinese requests such as skill 语言规范, Skill 中英文怎么分层, 哪些字段必须英文, 触发短语是否该保留中文, or 检查 skill 是否过度英文化.

Use when you need to execute a written implementation plan in the current conversation by assigning mostly independent tasks to fresh implementation-capable subagents with two-stage reviews, including Chinese requests such as “每个任务派子代理做”“子代理驱动开发”“按任务拆给子代理实现”. Do not use when the environment lacks real implementation-capable subagents, the tasks are tightly coupled, or executing-plans is the safer path.

Troubleshoot Chrome DevTools MCP connection, startup, target selection, navigation, and missing-tool issues in Trae on Windows. Use when Chrome DevTools MCP tools such as list_pages, new_page, navigate_page, take_snapshot, click, lighthouse_audit, or performance tools fail, or for Chinese requests such as MCP 启动失败, 浏览器连不上, 工具缺失, 页面打不开, or Chrome DevTools 报错.

Provide a browser-based prototyping companion for brainstorming sessions — show layouts, wireframes, architecture diagrams, and visual options, and collect the user's click choices. Use when the user has accepted the visual companion offer during brainstorming, or when you need to display visual content for user feedback, including Chinese requests such as “在浏览器里看原型”“展示布局对比”“让用户点击选择”. Requires node available and a local port accessible. Do not use for pure text discussions, concept clarification, or decisions that don't benefit from visual presentation.