开发流程五步法。支持需求理解、方案设计、代码实现、代码审查、Bug 修复。当用户提到「需求分析」「方案设计」「代码实现」「代码审查」「理解需求」「技术设计」「开始写代码」「Review」「检查代码」「bug」「报错」「崩溃」「异常」「出错了」时使用。

Codex-compatible Ralph loop runner with dual engines (compat local state loop + optional open-ralph-wiggum backend).

Generate changeset files for versioning and changelog management in this monorepo.

Real-time search engine supporting web search, vertical domain search, parallel batch search, and URL content extraction.

万悟平台 SSE 子会话递归嵌套与三明治序列渲染架构指南。涵盖 parentId 领养、order 绝对排序、动静 Chunk 分层及 Vue 2 响应式引用协议。

TypeScript code shape inside a well-named module — taste, not lint. Prefer one class or namespace per file (the unit a test targets) over scattered free exports; consolidate related code, don't fragment. The unit of code should be the unit of spec. Use when authoring TS in `editor/grida-canvas*`, `editor/lib/`, or `packages/*`. Sibling to the `naming` skill; React-specific shape lives in `code-react`.

Forms a GitHub search query based on a natural language query and the type of search (issue or PR). This skill helps users create effective search queries to find relevant issues or pull requests on GitHub.

Used for command-shape or live NV-Reason-CXR chest X-ray reasoning smoke tests. Not for diagnosis or clinical reporting.

Stage 3 of Clinical ASR Flywheel. Score a NeMo manifest, produce the five-section KER leaderboard (by-ipa_source diagnostic). Not for ASR auth (/riva-asr).

Audit changed Android translations in a GitHub PR.

Replace the usage of `expect` and similar functions from `package:matcher` to `package:checks` equivalents.

Diagnose SQLiteCpp build, linker, compiler, and sqlite3 integration issues.

Default response behavior for AI coding agents: professional, factual, neutral tone with calibrated critical evaluation. Baseline for every task: implementation, code review, debugging, planning, research, evaluation, Q&A.

Add and manage evaluation results in Hugging Face model cards. Supports extracting eval tables from README content, importing scores from Artificial Analysis API, and running custom model evaluations with vLLM/lighteval. Works with the model-index metadata format.

Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training

Fine-tune LLMs on Google Colab GPUs directly from openscience. Connects to Colab runtimes via WebSocket bridge for remote training with Unsloth. Supports SFT, GRPO, DPO, vision, and TTS workflows on free T4 to Pro A100 GPUs.

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques. Use when optimizing inference speed (1.5-3.6× speedup), reducing latency for real-time applications, or deploying models with limited compute. Covers draft models, tree-based attention, Jacobi iteration, parallel token generation, and production deployment strategies.

Creates, reviews, and dispatches atomic work packets. Every task with a side effect requires an APPROVED packet before execution. Scope is immutable after approval — expansion requires a new packet.

Use when Codex needs to understand or respond to automatic comment-checker feedback emitted after an edit-like PostToolUse hook.

End-to-end validation of coordinator and agent template changes