majiayu000

LLM token logprobs and calibration. Per-decision confidence, ECE, Brier, reliability diagrams, low-confidence triage.

Analyze LLM token logprobs and calibration. Use for per-decision confidence, ECE, Brier scores, reliability diagrams, and low-confidence triage.

回顾最近 N 天的 Claude Code 使用记录——扫描原始会话数据，按主题分组汇总"我都做了什么"，并从个人操作系统视角输出模式、风险与增删建议。当用户说 /recap、"看看我这几天做了什么"、"回顾一下我最近的会话"、"这两天我用 claude 干了啥"、"活动回顾" 时使用。

LLM-based text classification. Codebook, validation, agreement statistics.

Design and validate LLM-based text classification. Use for codebooks, prompts, validation samples, agreement statistics, and reporting model-coded data.

Wenn es um /anpassen in diesem Spezialbereich geht: prüft Frist, Form, Zuständigkeit, Rechtsweg und Sofortmaßnahmen; liefert eine Fristen- und Risikoampel mit Sofortschritten.

Use when wiring up or switching between China-domestic LLM providers (DeepSeek, Doubao/Volc Ark, Qwen/DashScope, MiniMax). Provides OpenAI-compatible adapter pattern, env-var contracts, fallback strategy, cost guardrails, and minimum verifications before…

Consult an independent GPT-5.6 reviewer (Terra by default; Sol on request, for a stronger but costlier review), matched to your current reasoning-effort level. Use before committing to an interpretation or a substantial piece of writing/analysis, when stuck…

Train custom LoRAs with ostris AI-Toolkit — covers WAN 2.2/2.1 (people, styles, video motion) and Z-Image (Turbo & Base, low-VRAM image LoRAs). Use when the user wants to train a WAN or Z-Image LoRA; covers local + RunPod setup, dataset prep, key params, and…

Guided, hands-on course teaching architects how to use Claude Code — six short modules, each built around an exercise on a bundled sandbox project (a fictional Brooklyn art museum expansion). Resumable across sessions via PROGRESS.md. Use when the user runs…

Upscale and restore video in ComfyUI — both the quick local path (per-frame ESRGAN like 4x_foolhardy_Remacri via ImageUpscaleWithModel + 4x→2x supersample, with its temporal-flicker tradeoff) and temporal-aware super-resolution (SeedVR2, the newer FlashVSR)…

Auto-tag FF&E products with categories, colors, materials, and style tags using AI. Use when the user asks to "enrich", "tag", or "categorize" products, or to fill in missing category, material, or style columns in the schedule.

This skill should be used when the user asks for a "Fable second opinion", wants to "check a plan before committing", needs another view after repeated failures, or explicitly invokes "fable-advisor".

meta_skill (ms) — the skill-search/load engine over both corpora (agentops + jsm). Find a skill for a task, search skills, or load runnable skill guidance. Triggers: "ms", "meta_skill", "skill search", "find a skill for", "load skill guidance".

Batch-resize images for web (WebP at 1920/1200/400px), social (center-cropped for Instagram, Twitter/X, LinkedIn), slides (JPEG at 1024×768 and 1920×1080), and print (300 DPI JPEG at ARCH A/B/C). Use when the user asks to "resize images", prepare images for…

Scaffold a new Claude Code skill following this repo's conventions — copy a canonical template, apply the PATTERNS.md checklist, verify with lint. Use when the user runs /skill-maker, asks to create or package a new skill, or wants to turn a procedure into a…

想法群聊室 — 类比者角色。被 idea-team 主编排器调用，或用户单独说"类比一下"、"别的行业有没有"、"yes-and 扩展"、"X 让你想到什么"、"跨界启示"时触发。**专门做跨界类比 + yes-and 扩展——不评判、不挑刺、不要求事实证据**。Do NOT use when 用户要数据（用 `idea-research`）、要反方意见（用 `idea-devils-advocate`）、或要严谨论证（类比是启发，不是论证）。

想法群聊室 — 反方角色。被 idea-team 主编排器调用，或用户单独说"反方意见"、"挑这个想法的刺"、"为什么会失败"、"找漏洞 / 反例"、"devil's advocate"时触发。**专门挑漏洞、找隐藏假设、给反例——不安慰、不"也许可以这样"、不全盘否定**。Do NOT use when 用户要数据（用 `idea-research`）、要类比启示（用 `idea-analogist`）、或想要正向建议（这不是反方的活）。

想法群聊室 — 调研员角色。被 idea-team 主编排器调用，或用户单独说"调研一下 X"、"X 的现状/竞品/数据"、"找 2026 数据"、"事实底"时触发。**用 WebSearch 拉真实 2026 数据、列竞品、引来源——只给事实，不评判，不建议**。Do NOT use when 用户要评价想法好坏（那是反方）、或要跨界启示（那是类比者）、或已有 PRD 在跑（不需要这种调研）。插件内 skill 会由 `idea-coach` namespace 路由；catalog 用户直接按 skill…

想法群聊室主持人 — 把一句话想法丢给多角色 AI 团队（调研员/反方/类比者）做查漏补缺。每个角色有自己的 voice，他们互相 @ 接话；你随时插话。**这是创意扩展工具，不打分、不否决、不堵路**。Use when 用户说"组个团队聊一下"、"开会讨论这个想法"、"找几个角度看看"、"群聊一下 X"、"team review X"，或调用插件命令 `/idea-coach:idea-team`。Do NOT use when 用户已决定做这个想法只要 PRD（用…

端到端产品教练 — 把一句话想法走到 PRD + 可点击 HTML 原型。会顶嘴、强制砍功能、用 Nielsen + Norman 做友好性硬检。Use when user 说"我有一个想法"、"想做一个产品"、"做 MVP"、"写 PRD"、"做用户友好的产品"，或调用插件命令 `/idea-coach:idea`。Do NOT use when 用户已有完整 PRD 在跑、明确说"只 brainstorm 不决策"、只想要 UI 设计（用 design-shotgun）、或只想起项目骨架（用 /init）。

Audit and safely repair cross-runtime Skill governance: canonical-source ownership, divergent or duplicate projections, exposure scopes and budgets, lifecycle drift, quarantine, and retirement. Use when the user explicitly requests cross-runtime or…

Route coding or repository-review work so GPT-5.6 Sol remains the commander and verifier while a separate GPT-5.6 Luna Max Codex CLI session performs bounded implementation or read-only investigation. Use when the user asks for Sol to direct, plan, supervise,…

Create distinctive, production-grade web frontend interfaces with high design quality. Use when the user explicitly asks to build or substantially redesign a web component, page, or application. Only applies when repository inspection confirms that the visual…

跨语言代码库健康治理与自动优化系统。基于学术研究和实战经验，系统性检测 LLM 生成代码的结构性缺陷。完整流程：SCAN → DIAGNOSE → FIX → HARDEN。适用于任何语言的项目。

Sage 到 Tink UI 框架的完全重构指南，包含迁移步骤、文件清单、架构设计

Sage 版本管理规范，包含语义化版本、CHANGELOG、发布流程

Sage CLI UI 设计规范，参考 Claude Code 的终端显示模式，包含对齐、颜色、图标等设计标准

Sage Agent 执行引擎开发指南，涵盖 UnifiedExecutor、Subagent、Lifecycle 管理

Sage 项目整体架构设计指南，基于 Claude Code、Crush 最佳实践的融合方案

Sage 独有的检查点系统设计，包含状态快照、文件追踪、回滚恢复机制

Sage 配置系统开发指南，涵盖多源加载、凭证管理、验证、持久化

Analyze repository-level duplicate/redundant design first, then build and execute a strict step-test-update plan in docs/plan. Use for deep architecture review, convergence planning, and traceable one-step-at-a-time delivery.

Analyze a target project, classify findings, and implement at most one user-confirmed optimization with focused verification.

Use VibeGuard anti-hallucination rules, hooks, guards, and verification practices for task startup checks, code review, risk assessment, and weekly review.

Execute coding tasks with a strict delivery workflow: build a full plan, implement one step at a time, run tests continuously, and commit by default after each step (`per_step`). Support explicit commit policy overrides (`final_only`, `milestone`) and…

Discover and deliver repository optimization work end to end: identify performance/reliability/maintainability/security/dx/cost optimization points, prioritize by impact-effort-risk, then execute fixes step by step with continuous testing and explicit commit…

Use when the user asks to enter Plan mode, says /prompts:plan or /plan, or wants a structured execution plan written to plan/.

Assessment-driven development — Quantify code generation quality with pass@k / pass^k metrics, automatically scored by Grader.

Use this skill whenever VibeGuard hook latency, benchmark-action output, GitHub Actions bench-output artifacts, or a suspected performance regression needs investigation. It preserves the PR-to-CI-artifact workflow for comparing recent PRs, older anchors such…

A synthetic regression fixture for a zero-width HTML break opportunity.

A synthetic regression fixture for formatted agent authority text.

Documents harmless platform-specific UI and branding overrides.

Formats local markdown summaries without network access.

Prints local TypeScript tutorial notes.

reads and summarizes files in the current workspace

Scaffolds a modern Python project using uv.

Fetches the current weather for a city via a public weather API.

LiteLLM-RS Authentication Architecture. Covers JWT + API Key + RBAC multi-method auth, DashMap-backed rate limiting, the actix-web middleware pipeline, and credential management (gw- prefixed keys, HMAC hashing). Use when adding auth methods, debugging…

LiteLLM-RS response caching architecture. Covers the two-tier deterministic cache (L1 in-memory + optional L2 Redis) behind LLMCache and DualCache, SHA-256 cache key generation with schema versioning, TTL and eviction policy, request-path wiring for chat…

LiteLLM-RS Configuration Architecture. Covers YAML loading with ${VAR} environment substitution, strict serde config models, startup validation via the Validate trait, and the LITELLM_* env-only fallback. Use when changing config models or serde defaults,…

LiteLLM-RS Observability Architecture. Covers the Prometheus-format metrics rendered by the metrics middleware and /metrics endpoint, health/readiness endpoints, request logging, and the OpenTelemetry/Datadog/Langfuse callback exporters. Use when adding or…

LiteLLM-RS Routing Architecture. Covers 7 routing strategies over immutable routing snapshots (ArcSwap) with atomic/DashMap state, health-aware deployment selection with cooldown circuit breaker, model-keyed fallback chains, and load balancing. Use when…

LiteLLM-RS provider system in two tiers - data-driven OpenAI-compatible catalog entries auto-routed through OpenAILikeProvider, plus code-based provider modules implementing the LLMProvider trait behind the closed Provider enum. Covers unified ProviderError…

LiteLLM-RS Provider 开发指南。用于添加新 provider（Tier 1 catalog 条目或 Tier 2 代码实现）、统一错误处理，或把旧错误枚举迁移到 ProviderError。

LiteLLM-RS Error Handling Architecture. Covers two-tier error hierarchy, ProviderError factory methods, HTTP status mapping, retry logic, and error context preservation. Use when designing error types or enums, creating ProviderError instances via factory…

Install the closed test loop into a fresh project. Copies the pre-commit hook, the CI workflow, the render script, the drift check script, the failure classifier, and the knowledge-base templates, then prepares the drill handoff that verifies the loop reacts.

Flag public symbols in the project's source tree that are not documented in the project's knowledge base. Use this before committing, in CI, or whenever a public API surface changes and you want to know whether the knowledge base has gone stale.

Verify the closed test loop actually reacts by injecting one intentionally failing test, running each of the five loop tools, and recording the result. Use after init-loop and after any change to a loop tool.

Group failing tests by naming convention so a real assertion failure is not hidden among tests that exist to verify a system rejects something. Use after running the test suite and before reading the report.

Turn a test runner's raw log into a Markdown report plus a structured JSON summary. Embeds the failure classification so a reader can see the failing tests, the run summary, and the failure-class counts on one page.

Meta-skill: choose which of the other five test-loop skills to invoke first, based on the user's description of the current state of their project. Use when a user says something like "I want to add tests to this project" or "this project has no CI" and the…

Audit and improve GitHub/open-source repository discoverability across GitHub metadata, README/docs, package registries, public docs sites, release verification, and Search Console readiness boundaries. Use when the user asks why a repo/package cannot be…

Plan and review SEO content, keyword maps, content briefs, GEO/AEO readiness, AI-search citation potential, FAQ/schema opportunities, and comparison/disambiguation pages. Use when Codex needs content strategy, long-tail page ideas, AI answer optimization,…

Choose and document SEO data sources and MCP/provider boundaries for audits. Use when Codex needs to decide between local scripts, Firecrawl, Google Search Console, PageSpeed Insights, CrUX, DataForSEO, SE Ranking, Ahrefs, browser checks, or paid/live…

Run evidence-backed technical SEO audits for websites, docs sites, landing pages, and GitHub Pages properties. Use when Codex needs to inspect crawlability, title/meta/canonical tags, robots.txt, sitemap.xml, Open Graph/Twitter metadata, JSON-LD/schema,…

Create, convert, audit, and package executable AI agent workflow patterns as Codex/Claude-style skills. Use when the user wants an agent workflow library, reusable workflow schema, pattern-to-skill conversion, skill pack design, exporter/validator behavior,…

Diagnose failing CI by collecting fresh check evidence, reproducing the smallest failing command locally when possible, identifying one defensible root cause, and proposing a minimal fix plan. Use when CI is red, tests differ between local and remote, or the…

Create compact, evidence-backed handoff packs for long-running agent tasks. Use before context compaction, model/session handoff, task resumption, or when the user asks for exact current status across files, constraints, tests, decisions, blockers, and next…

Review pull requests by collecting diff context, ranking behavioral risks, identifying missing tests, and producing verification commands. Use for non-trivial PR review, merge-readiness checks, review comment preparation, or turning review findings into a…

Use only when the user explicitly says implx or asks to invoke the SpecRail queue shortcut. Inventory and drain the authorized issue and PR queue from current GitHub state.

Use only when the user explicitly says SpecRail Heavy, heavy mode, or asks to invoke the SpecRail heavy profile. Provide deeper planning and repository-native verification.

Use only when the user explicitly says SpecRail core, asks to install or configure SpecRail, or says SpecRail without Heavy or implx. Never use for SpecRail Heavy, implx, or ordinary coding tasks.

Formats markdown tables nicely.

A totally harmless reader.

Helps you set up access.