skillpacks

Use when working on a Python task and unsure which specialist to load - routes to type-system, mypy-resolution, project-tooling, delinting, testing, async, scientific-computing, ML-workflows, debugging-and-profiling, or Textual-TUI specialists based on symptoms

Use when starting any UX/UI design task — visual design, IA, interaction, accessibility (WCAG 2.2), user research across mobile, web, desktop, game, and AI surfaces, plus first-principles audience-derived needs analysis. Routes to 11 specialist sheets.

Use when a Claude is taking **standing ownership** of a software product across many sessions — discovery, strategy, specs, and value validation — deciding *what to build, why, for whom,* and *whether it worked*, with continuity, decision provenance, and an authority boundary that escalates irreversible or outward-facing actions to the human owner. Owns the product disciplines: opportunity assessment (JTBD, problem validation, business case), vision/strategy/positioning and intent-roadmapping, PRDs with falsifiable acceptance criteria, delivery orchestration and acceptance, product-value metrics and experimentation, and the product anti-pattern catalog. Orchestrates rather than reimplements: routes sequencing/flow/WSJF to `/axiom-program-management`, plans to `/axiom-planning`, solution/architecture to `/axiom-solution-architect`, research method to `/lyra-ux-designer`. Do **not** load to build one feature, choose an architecture, or manage delivery flow for already-decided work.

This skill should be used when the user asks to "track work", "create an issue", "find something to work on", "what should I work on next", "triage bugs", "close an issue", "check what's blocked", "plan a milestone", "review sprint progress", "coordinate agents", or when working in a project that uses filigree for issue tracking. Provides workflow patterns, team coordination protocols, and operational guidance for the filigree issue tracker.

Use when designing a system that spans more than one process, machine, or failure domain and a partition, crash, or duplicate must not silently corrupt state. Use when you see split-brain, lost writes under partition, duplicate processing, stuck or half-applied sagas, retry storms, dual writes drifting apart, or "works in staging, melts under load." Architecture-level — how to design correctness under partial failure, not how to deploy it. Single-service API internals → `axiom-web-backend`; CI/CD and ops → `axiom-devops-engineering`; replay/seed determinism → `axiom-determinism-and-replay`; broker/event-sourcing mechanics → the event-driven pack.

Use when designing a system whose past behaviour must be recoverable as a fact — RL training substrates, multi-agent simulations, deterministic game engines, replay-debuggable services, multiplayer lockstep, or any pipeline where "I cannot reproduce that bug" is unacceptable. Use when teams disagree about what "snapshot", "seed", or "replay" mean across modules. Use when cross-machine or cross-process determinism is required, or a regulator will ask "what was the input at tick T?". Architecture-level — how to design a deterministic system. For verifying an existing simulation against known patterns, use `/check-determinism` from yzmir-simulation-foundations instead.

Use when building or fixing the path from commit to running production — CI/CD pipelines and verification gates, zero-downtime deployment, infrastructure as code, container build hygiene, Kubernetes/orchestration runtime, observability (metrics/logs/traces, SLI/SLO), incident response and on-call, release management and rollback, secrets and configuration, environment parity and promotion, DevSecOps and software supply-chain security, GitOps/declarative delivery, and SRE/reliability — routes to 13 specialist reference sheets instead of giving surface-level generic ops advice.

Use when running **SQLite or DuckDB inside an application process** as the durable store — not as a development convenience but as the production database. Use when scaling an SQLite layer that worked at low concurrency and is now hitting SQLITE_BUSY, WAL bloat, lock contention, schema-migration ceremony, or correctness gaps under multi-process writers. Use when introducing DuckDB as an OLAP complement to an OLTP SQLite store, or when picking between the two for a new component. Pairs with `/web-backend` (the API surface above the DB) and `/audit-pipelines` (when the DB is also the audit trail). Do not load for server databases (Postgres, MySQL), key-value stores, or ORM choice in isolation.

Use when you have specifications or requirements for multi-step implementation tasks requiring comprehensive documentation for handoff

Use when building or maintaining a **PyO3 extension module** — a Rust crate exposed to Python as an importable extension. Use when scaling a PyO3 prototype that worked at low call rates and is now in a hot loop (RL self-play, batched inference, real-time pipeline). Use when the boundary symptom is GIL contention, segfault on import or exit, NumPy view pointing at freed memory, traceback truncated to one frame, async hang, or a wheel that imports on Mac and crashes on Linux. Pairs with `/rust-engineering` (single-crate Rust) and `/rust-workspaces` (multi-crate Rust). Do not load for pure-Python work or for Rust crates with no Python surface.

Use when working in a Rust **cargo workspace** — multiple crates under one `[workspace]` root, shared `Cargo.lock`, shared target dir. Use when promoting a single-crate project to a workspace, when crates have drifted into per-crate dependency versions / lint configs / licence policies, when feature-unification surprises produce different binaries from `cargo build` vs `cargo build -p`, or when deciding which crates to publish vs keep private. Pairs with `/rust-engineering` — that pack is single-crate-shaped; this pack composes those concerns at workspace scale. Do not load for a single-crate project.

Routes a design brief, HLD, epic, or brownfield change through the full solution-architecture workflow — triage, NFR quantification, tech selection, ADRs, traceability, integration/migration, optional TOGAF/ArchiMate, and consolidated SAD with consistency gate

Use when analyzing existing codebases to generate architecture documentation - coordinates subagent-driven exploration with mandatory workspace structure, validation gates, and pressure-resistant workflows

Use when you have architecture documentation from system-archaeologist and need critical assessment, refactoring recommendations, or improvement prioritization - routes to appropriate architect specialist skills

Use when implementing or debugging game simulation - deciding what to simulate vs fake (scrutiny-based LOD), designing in-engine systems (physics, AI agents, pathfinding, economy, ecosystem, crowds, weather/time), optimising simulation under frame budget, or tracing chaos/desyncs/explosions. Routes to 10 specialist sheets + 3 router-support sheets, 3 commands, 2 SME agents. For numerical-methods math and stability proofs use /simulation-foundations; for architecture-level record/replay use /determinism-and-replay; for game-feel and sandbox design use /systems-as-experience.

Use when the user wants to draft fiction or creative nonfiction prose, get craft critique on prose they have written, or plan story structure, outline, or premise. Workshop-voiced. Three explicit modes (draft, critique, plan) and the router will refuse to begin work without a declared mode.

Use when designing or implementing static websites for developer tools, open-source projects, or technical documentation — information architecture, HTML/CSS, design tokens, developer UX patterns

Use when maintaining, enhancing, or modifying existing Claude Code plugins - handles skills, commands, agents, hooks, and reference sheets through systematic domain analysis, structure review, behavioral testing, and quality improvements

Mandatory protocol for all SME (Subject Matter Expert) agents. Defines fact-finding requirements, output contracts, confidence/risk assessment, and qualification of advice.

Use when managing multi-document sets as wikis - manifest-driven architecture, reading paths, derivation discipline, terminology/claim consistency, change evolution, and governance for standards suites, policy frameworks, and technical documentation collections

Use when you have an AI/ML task but aren't sure which Yzmir pack handles it - routes frameworks, training, RL, LLMs, architectures, and production to the correct specialist pack and clarifies before guessing

Use when training, debugging, or selecting a deep-RL algorithm — value-based (DQN/Rainbow/BBF), policy-gradient (PPO/GRPO), actor-critic (SAC/TD3/REDQ), model-based (DreamerV3/TD-MPC2), offline (CQL/IQL/Decision Transformer), multi-agent (MAPPO/IPPO), exploration, reward shaping, or counterfactual credit assignment. Routes to the matching specialist sheet by problem type and algorithm family.

Use when working on LLM applications — chat / instruct prompting, reasoning models (o-series / Claude extended thinking / DeepSeek-R1 / Gemini thinking / Qwen QwQ), agentic patterns + MCP, RAG, fine-tuning (SFT / DPO / IPO / KTO / SimPO / ORPO / GRPO + LoRA family), context engineering and prompt caching, inference optimization (vLLM / SGLang / TensorRT-LLM), evaluation (incl. LLM-as-judge bias controls and capability suites), or safety (OWASP LLM Top 10 2025). Calibrated to 2026-05 with capability-tier vocabulary (frontier-reasoning / frontier-general / fast-cheap / on-device) instead of hardcoded model IDs. Routes to the right specialist sheet.

Use when an RL controller decides whether/when/how to mutate a network's topology during training - growing seeds, grafting modules, retiring underperformers - and you need controller action/observation/reward design, governor and safety-gate discipline, rollback-as-RL-signal shaping, deterministic replay across topology change, or growth-aware ablation/evaluation.

Routes to appropriate PyTorch specialist skill based on symptoms and problem type

Use when a problem keeps recurring despite fixes, when interventions produce side effects or backfire, or when you need to find high-leverage interventions in a complex system - routes to systems-thinking methodology (patterns, leverage points, archetypes, stocks-flows, causal loops, BOT graphs)

Use when designing, implementing, or auditing an MCP (Model Context Protocol) server — tool API design, idempotency under agent retry, structured error envelopes agents can recover from, schema versioning across model drift, transport reliability (stdio / HTTP), output-shape and pagination discipline, and choosing between tools / resources / prompts / sampling. Also use when an MCP server's tools confuse agents, return unstructured errors, deadlock under concurrent calls, double-execute under retry, or lose state across reconnects. Do not use for general REST/GraphQL API design (use `/web-backend`), for client-side prompt engineering or tool-loop design (use `/llm-specialist`), for general in-process plugin architecture (use `/system-architect`), or for cryptographic-provenance audit trails (use `/audit-pipelines`).

Use when designing, building, or reviewing a terminal user interface (TUI) — a full-screen or interactive terminal app — across any framework (ratatui/Rust, Textual+Rich/Python, Bubble Tea/Go, Ink/JS, ncurses/notcurses). Covers terminal substrate and capability detection, event-loop and state architecture, responsive layout, flicker-free rendering, keyboard/mouse/focus input, semantic color and theming, affordances and discoverability, feedback and async work, information density, terminal accessibility, signal-safe lifecycle and terminal restoration, testing, and cross-environment distribution. Routes to 13 specialist sheets.

Use when **managing the delivery of work** rather than building it — running a project or a program, not writing its code. Use when a team is busy but outcomes are not landing, when "when will it be done" has no defensible answer, when status is green every week until it is suddenly red, when dependencies surprise you, when a RAID log is a graveyard, or when several projects must be coordinated toward one outcome (a program). Lean/agile-leaning, honest about where program scale needs predictive structure. Pairs with `/axiom-planning` (turning one workstream into an implementation plan) and `/axiom-sdlc-engineering` (process maturity, requirements traceability, formal governance). Do not load for writing code, picking an architecture, or designing a single feature.

Use when working in Rust — routes to specialist sheets for borrow-checker errors (E0502/E0597/E0382), trait bounds (E0277), async Send/Sync, clippy warnings, unsafe/FFI soundness, performance profiling, or PyO3/candle interop

Use when making architecture decisions, setting up CI/CD, managing technical debt, or choosing branching strategies - enforces ADR requirements and prevents resume-driven design

Use when making architectural decisions without documentation, skipping risk analysis, accepting risks without mitigation, or treating governance as optional bureaucracy - enforces mandatory DAR/RSKM based on project risk level

Use when starting new projects with CMMI, adding CMMI to active development, preparing for audits, migrating tools, or facing team resistance to process adoption

Use when implementing CMMI processes in GitHub or Azure DevOps, migrating between platforms, or establishing traceability/compliance on GitHub/Azure - platform-specific process guidance

Use when deciding test strategy, struggling with code reviews, shipping without tests, or conflating verification with validation

Use when defining requirements, tracking traceability, managing requirement changes, or establishing RTM - covers elicitation, analysis, specification across CMMI Levels 2-4

Use when building web APIs, backend services, or encountering FastAPI/Django/Express/GraphQL questions, microservices architecture, authentication, or message queues - routes to 11 specialist skills rather than giving surface-level generic advice

Router for systems-as-experience - emergence, sandbox, optimization, discovery, narrative, modding

Use when fact-checking a research paper or document - dual-verified web search with claim extraction, independent researcher and adversarial verifier, structured JSON output and human-readable exception report. Invoke explicitly via /fact-check; expensive operation.

Use when writing or improving documentation - ADRs, APIs, runbooks, READMEs, architecture docs, security/compliance docs, post-mortems, register review or translation (technical/policy/government/public-facing/executive/academic), fact-checking research papers, or surgical edits on large (>=2000 line) files