ANcpLua

Publish a Skill, config, or plugin to TomeVault as a high-grade Tome — and understand the Skill / Tome / AGENTS.md model behind it. Use when shipping instruction files to TomeVault, indexing a repo, aiming for a Gold grade, passing the TomeVault scan or validator, or answering "what is a Tome / Skill / AGENTS.md". Covers the npx tomevault CLI, the Studio web publish flow, and the grading rubric.

Ferry a GitHub PR to merge — one snapshot-not-block iteration that reads live PR state, classifies it, and runs exactly one handler: fix CI at the root cause, repair merge conflicts, triage reviewer threads with version-currency checks, rewire human-blocked reviews, or propose-and-pause before any force op. Use when asked to merge / land / babysit a GitHub PR — including plain-language 'merge my PR', 'get this branch merged', or 'babysit my PRs' with no slash command and no state file yet — or when resuming a /charon ferry, without ever waiting forever.

Use when the user needs X/Twitter data through Xquik for REST API planning, MCP setup, SDKs, search, exports, monitoring, webhooks, or confirmation-gated publishing.

Analyze CI build and test status from Azure DevOps and Helix for dotnet repository PRs. Use when checking CI status, investigating failures, determining if a PR is ready to merge, or given URLs containing dev.azure.com or helix.dot.net. Also use when asked "why is CI red", "test failures", "retry CI", "rerun tests", "is CI green", "build failed", "checks failing", or "flaky tests". DO NOT USE FOR: investigating stale codeflow PRs or dependency update health, tracing whether a commit has flowed from one repo to another, reviewing code changes for correctness or style.

Run risky changes through 8 disciplined phases that map, build, fix, and verify — each unlocks more power only if work stays honest and resumable. Use when a spec or audit needs parallel discovery, lane execution, checkpoints surviving context loss, and a ship-or-die finale.

Analyze, lint, and check a local Claude Code add-on or capability, explain the result, and rank the top fixes first. Use when the user says "audit this", "why did it score that way", "what should I fix first", "validate the manifest", "lint the hooks", or "benchmark the real token usage".

Audit a local Claude Code bundle that has .claude-plugin/plugin.json: validate the manifest, lint hooks, check mcp/lsp servers, monitors, and agents, then rank fixes. Use when the user says "audit this", "validate the manifest", "lint the hooks", "why did it score that way", or "benchmark this".

Analyze and check a local Claude Code SKILL.md folder: review frontmatter, size, broken links, trigger words, and token budget, then rank the top fixes. Use when the user says "audit this", "why did it score that way", "this is too big", "it is not triggering", or "measure the real token usage".

Write Microsoft Agent Framework (.NET) code against the cloned, SHA-pinned source — never from memory or from Microsoft Learn, which lag the source by weeks and keep pre-GA signatures alive long after they were renamed. USE FOR: writing or reviewing any code that touches Microsoft.Agents.AI / Microsoft.Agents.AI.Abstractions (AIAgent, ChatClientAgent, AgentSession, AgentResponse, RunAsync/RunStreamingAsync) or Microsoft.Extensions.AI (IChatClient, ChatMessage, ChatRole, ChatOptions, AITool); a compile error or "method not found" on a MAF type; porting a doc/blog sample that uses AgentThread / AgentRunResponse / CompleteAsync / GetNewThread; deciding whether to wrap an IChatClient in an agent vs hand-rolling the tool loop. ESPECIALLY use it the moment you are about to emit a MAF type or method name from memory — that is exactly when the stale-rename traps bite. DO NOT USE FOR: Microsoft.Extensions.AI internals unrelated to agents; non-.NET MAF (Python); conceptual "why" questions where Learn is fine (use micro

Authoritative reference for the Model Context Protocol C#/.NET SDK 1.4.1. Use for MCP servers, clients, tools, prompts, resources, transports, sessions, tasks, sampling, elicitation, roots, identity, auth, filters, completions, logging, pagination, HTTP context, McpServer, McpClient, and ModelContextProtocol.* APIs. Always load the relevant reference file before answering.

Enable, wire up, and operate automated browser-extension publishing to all three stores — Microsoft Edge Add-ons (Partner Center Publish API, ApiKey auth), Chrome Web Store (API v2, Google OAuth refresh token), and Firefox Add-ons / AMO (API v5, HS256 JWT). Use whenever the task touches publishing or updating a browser extension on any store, a publish:edge / publish:chrome / publish:firefox script, the env vars EDGE_PRODUCT_ID / EDGE_API_KEY / EDGE_CLIENT_ID / CWS_CLIENT_ID / CWS_CLIENT_SECRET / CWS_REFRESH_TOKEN / CWS_PUBLISHER_ID / CWS_ITEM_ID / AMO_JWT_ISSUER / AMO_JWT_SECRET, enabling a store publish API, a failing upload, "submission in progress" / "pending review" errors, or automating extension releases — and ESPECIALLY when you are tempted to conclude publishing is blocked because API credentials are missing. Missing credentials are a minutes-long console fix, never a blocker.

Use when the user wants help writing, refining, debugging, or evaluating a prompt, system prompt, custom instruction, agent prompt, or skill description. Triggers on phrases like "improve this prompt", "why isn't Claude doing X", "write a system prompt for...", "review my instructions", "this prompt isn't working", "the model keeps doing Y", "how should I phrase this", or any request to design, fix, or analyze AI prompts. Also triggers when the user is iterating on Claude Code skills, agent instructions, CLAUDE.md / AGENTS.md files, description fields for triggering accuracy, few-shot example design, or debugging unexpected model output. Use this even when the user hasn't named prompt engineering explicitly — if they're discussing a prompt's behavior or how to phrase something for a model, this applies. Also use for Claude Fable 5 / Mythos 5 prompting — migrating prompts from older models, effort tuning, agentic behavior issues, or unexpected refusals.

Weave Microsoft-Learn-grounded research into ANY substantive help for this user, whose work is overwhelmingly Microsoft-centric (qyl on .NET / Foundry / Microsoft Agent Framework; the microsoft-ai-hackathon tracks Otto, Agent Battle Arena, Inventory Reorder Copilot). Reach for this at the START of essentially any research, design, debugging, "how do I", "is this current / latest", API-shape, SDK-error, or build task — proactively consult the Microsoft Learn MCP and the microsoft-docs skills and the local ground-truth checkouts BEFORE answering from memory, even when the user never types "Microsoft", "Azure", ".NET", or "docs". This is a routing / insurance skill that exists specifically so Microsoft grounding stops getting skipped ~half the time. Stand down only when the task is explicitly about a non-Microsoft stack (AWS, GCP, OpenAI, etc.) or is a trivial mechanical edit with no knowledge question. Pairs with the microsoft-learn-grounding skill.

Operate the Microsoft Learn MCP — the free, no-auth dogfood API behind Ask Learn and Copilot for Azure — at expert level, and verify how fresh any Microsoft Learn page actually is. Use whenever grounding an answer in Microsoft / Azure / .NET / M365 / Foundry / Agent Framework / Copilot Studio docs, deciding whether a Learn page is current, looking up a page's "Last updated" date / ms.date / source commit, capping token spend on doc searches, or debugging why the Learn MCP returned nothing useful — even when the user never says "Microsoft Learn" or "MCP". Covers the three tools (microsoft_docs_search, microsoft_docs_fetch, microsoft_code_sample_search), the maxTokenBudget control, dynamic tool discovery, the daily-refresh freshness model, and the key gotcha that the MCP strips last-updated metadata so you must read it from the rendered page instead.

Run a maximally strict structural code review that hunts complexity-collapse opportunities, oversized files, creeping conditionals, and unearned abstractions, then defends every finding against refutation before reporting it. Scope the diff first, review against eight standards, and gate approval hard. Use for a supercritical review, deep maintainability audit, especially harsh code-quality review, or when the user asks for a thermo-nuclear-style review.

Know, without being told, which target framework every project in ~/RiderProjects/qyl-workspace compiles for and what that allows or forbids — the net10.0 baseline (full modern BCL, AOT/trim-aware) versus the small netstandard2.0 Roslyn island (source generators + analyzers + the multi-targeted OTel libs). USE FOR: editing or reviewing any C# in qyl, Qyl.AutoInstrumentation, qyl-dotnet-autoinstrumentation, Qyl.OpenTelemetry.SemanticConventions; choosing whether an API/language feature is legal in the file you are touching; a build error like CS0103/CS1061/CS8370/"is not available in netstandard2.0" inside a *.SourceGenerators / *.Analyzers / *.generators project; "can I use HashCode / System.Text.Json / ArgumentNullException.ThrowIfNull / Span here"; reviewing a PR that spans both a generator and its runtime consumer. ESPECIALLY use it before you assume a *.SourceGenerators project can call net10 BCL (it cannot — it is netstandard2.0), or that netstandard2.0 forbids records / init / required (it does NOT — th

Use for any request to create or redesign a UI, landing page, dashboard, hero, marketing page, or other visual frontend.

Use for any new feature, refactor, repository structure change, multi-file edit, or UI change. Produce the plan before code.

Use when working directly with Forgejo v15 APIs, private repositories, Actions workflows, runners, pull requests, reviews, statuses, releases, packages, webhooks, organizations, users, or admin operations. Prefer Forgejo's live Swagger and direct endpoints over GitHub-shaped assumptions.

Evidence-grounded technical-debt audit. Hunts accretion — code that was added (shimmed, wrapped, re-validated, try/catch-padded) when it should have been deleted — plus NativeAOT/source-generator hazards. Trigger with "tech debt", "technical debt audit", "what should we refactor", "code health", or when the user asks about code quality, refactoring priorities, or maintenance backlog.

Parse native Claude and Codex logs into normalized usage metrics. Use when analyzing token burn, cost per task, model mix, or budget tracking. Do not use for runtime behavior changes.

Parse native Claude and Codex logs into normalized usage metrics. Use when analyzing token burn, cost per task, model mix, or budget tracking. Do not use for runtime behavior changes.

Use when verifying the yt-transcript Chrome-only MV3 extension gate from this repository.

Provides comprehensive Tailwind CSS utility-first styling patterns including responsive design, layout utilities, flexbox, grid, spacing, typography, colors, and modern CSS best practices. Use when styling React/Vue/Svelte components, building responsive layouts, implementing design systems, or optimizing CSS workflow.

TDD workflow for FortiGate toolkit - guides test-first development with proper API mocking patterns