architect-stack-selector（架构师堆栈选择器）

Architect: Stack Selector Router

Use this skill when the user describes an app goal but has not chosen a concrete stack or skill set yet. This is the control plane skill: choose minimal compatible skills, apply them, and run validations.

Available Skills To Route

Base architect skills:

• architect-python-uv-batch
• architect-python-uv-fastapi-sqlalchemy
• architect-nextjs-bun-app
• architect-nextjs-vercel-app
• architect-next-prisma-bun-vector

Protocol architect skills:

• architect-nostr-intent-router

Add-on skills:

• addon-decision-justification-ledger
• addon-human-pr-review-gate
• addon-auth-access-control
• addon-observability-telemetry
• addon-observability-python-structlog
• addon-observability-nextjs-pino
• addon-domain-semantic-adaptation
• addon-deterministic-eval-suite
• addon-llm-judge-evals
• addon-rag-ingestion-pipeline
• addon-docling-legal-chunk-embed
• addon-nostr-client-nextjs
• addon-nostr-nip-profile-selector
• addon-nostr-nip-sync-lock
• addon-nostr-key-custody
• addon-nostr-nip23-longform
• addon-llm-ancient-greek-translation
• addon-llm-translation
• addon-direct-llm-sdk
• addon-langchain-llm
• addon-langgraph-agent
• addon-google-agent-dev-kit
• addon-vercel-ai-labs

UI skills:

• ui-social-feed-shell
• ui-editorial-writing-surface
• ui-document-intake-dropzone

Router Workflow

1. Parse user intent into:

• app surface: batch | api | web
• exposure model: public | internal (infer public by default for api and web unless the user clearly says internal-only)
• language/runtime preference: python | nextjs | unspecified
• data/retrieval needs: RAG yes/no, source formats (pdf/markdown/html/csv/text)
• protocol needs: Nostr yes/no
• Nostr protocol intent details: client mode, capability intents, explicit NIP constraints
• persistence needs: relational DB yes/no, vector DB yes/no
• domain language intent: preferred terminology, forbidden terms, tone, and user-facing naming constraints
• execution constraints: network access, available binaries (uv, bun, docker), sandbox restrictions
• delivery mode preference: production-default (default) or explicit NO_DOCKER=yes
• review gate preference: required (default) or explicit opt-out
• decision justification preference: required (default, non-optional)
• semantic adaptation preference: required (default) or explicit opt-out
• observability preference: required-for-public-api-web (default) or explicit opt-out
• observability intensity: baseline-runtime-default | explicit-full-telemetry-push (infer explicit-full-telemetry-push only when observability is a primary user requirement; otherwise keep the baseline default)
• eval preference: deterministic required (default), llm-judge optional (default)

1. Pick exactly one base architect skill.
2. Add zero or more add-ons.
3. Execute scaffold:

• Apply selected skills in order.
• Produce required artifacts for production-default mode.
• Run selected skills' validation checklists.

1. Return:

• selected skills (ordered)
• one-line reason per skill
• install commands
• short execution order
• execution mode (full or degraded fallback mode) when constraints are present
• whether containerization is in-scope (required by default unless user explicitly opts out)
• whether human PR review gate is in-scope (required by default unless explicitly opted out)
• whether decision justification ledger is in-scope (required by default, no opt-out)
• whether semantic adaptation is in-scope (required by default unless explicitly opted out)
• whether observability telemetry is in-scope (required for public api/web by default unless explicitly opted out)
• whether observability intensity is baseline runtime-default or explicit full telemetry push
• whether deterministic eval suite is in-scope (required by default unless explicitly opted out)
• whether llm-judge evals are in-scope (optional by default unless explicitly requested)
• Nostr protocol routing status and selected NIP profile (when applicable)
• scaffold status and validation results

Selection Rules

• If user asks for Python data processing or offline ingestion jobs: choose architect-python-uv-batch.
• If the user explicitly asks for Vercel deployment or a Vercel-optimized Next.js baseline: choose architect-nextjs-vercel-app.
• If environment is constrained (missing package managers and/or no network): keep selected architecture, but explicitly return a degraded fallback mode and what cannot be verified yet.
• Default to production-default. Docker is required by default across all runnable base architect skills.
• Default to addon-human-pr-review-gate.
• Default to addon-decision-justification-ledger.
• Default to addon-domain-semantic-adaptation.
• Default observability intensity to baseline-runtime-default.
• For baseline-runtime-default, add only the runtime-specific observability implementation add-on for public api and web surfaces:

- Python api or worker runtime: addon-observability-python-structlog - Next.js or TypeScript server runtime: addon-observability-nextjs-pino

• Default to addon-deterministic-eval-suite.
• Only allow local-no-docker when user explicitly sets NO_DOCKER=yes.
• Only allow skipping addon-human-pr-review-gate when user explicitly requests opt-out.
• Do not allow skipping addon-decision-justification-ledger in production-default mode.
• Only allow skipping addon-domain-semantic-adaptation when user explicitly requests opt-out.
• Only allow skipping observability add-ons for public api and web surfaces when user explicitly requests opt-out.
• Only allow skipping addon-deterministic-eval-suite when user explicitly requests opt-out.
• If user asks for qualitative scoring, rubric grading, or "LLM as judge": add addon-llm-judge-evals.
• If user asks for Python API with database/migrations: choose architect-python-uv-fastapi-sqlalchemy.
• If user asks for a Next.js app without heavy backend DB requirements: choose architect-nextjs-bun-app.
• If user asks for Next.js + Prisma + Postgres + vector retrieval in one app: choose architect-next-prisma-bun-vector.
• If user mentions RAG, retrieval, embeddings, or document QA: add addon-rag-ingestion-pipeline.
• If user mentions legal PDFs/contracts/clauses and wants doc-to-markdown chunking: add addon-docling-legal-chunk-embed (typically with addon-rag-ingestion-pipeline).
• If user mentions Nostr relays/events/signing: add addon-nostr-client-nextjs.
• If user mentions Nostr and has not provided a complete NIP profile: add architect-nostr-intent-router before Nostr add-ons.
• For Nostr projects by default: add addon-nostr-nip-profile-selector.
• If user asks to fetch/update current NIPs or maintain a NIP lockfile: add addon-nostr-nip-sync-lock.
• If user requires private key handling, signer custody, or non-extension signing fallback: add addon-nostr-key-custody.
• If user requires NIP-23 long-form posts or Kind 30023: add addon-nostr-nip23-longform.
• If user asks for LLM translation to Ancient Greek (koine/attic): add addon-llm-ancient-greek-translation.
• If user asks for general LLM translation/localization flows: add addon-llm-translation.
• If user asks for login, sessions, JWTs, API keys, RBAC, scopes, protected routes, or access control: add addon-auth-access-control.
• If user makes observability a primary requirement (for example asks for logging, metrics, tracing, health checks, production diagnostics, or telemetry as a named deliverable): set observability intensity to explicit-full-telemetry-push.
• If observability intensity is explicit-full-telemetry-push: add addon-observability-telemetry plus the runtime-specific observability implementation add-on when the runtime is known.
• If user asks for cross-runtime observability policy, asks for generic observability guidance before choosing a runtime, or the runtime does not have a dedicated implementation add-on: set observability intensity to explicit-full-telemetry-push; add addon-observability-telemetry.
• If user asks for direct provider SDK usage, explicit provider control, or to avoid LangChain/ADK/Vercel AI abstractions: add addon-direct-llm-sdk.
• If user asks for LangChain abstractions or LLM provider interactions (chat/completions/RAG chain execution): add addon-langchain-llm.
• If user asks for LangGraph orchestration or multi-step agent behavior (state machine, tool loops, checkpoints): add addon-langgraph-agent.
• If user explicitly asks for Google Agent Development Kit (ADK): add addon-google-agent-dev-kit.
• If user asks for Vercel AI SDK or AI Labs style chat/generation flows: add addon-vercel-ai-labs.
• If user asks for social timeline/feed UI (including X/Twitter-like structure): add ui-social-feed-shell.
• If user asks for minimalist/typography-first long-form writing UX: add ui-editorial-writing-surface (with Next.js + relevant add-ons).
• If user asks for PDF/Markdown upload intake UI: add ui-document-intake-dropzone (typically with addon-rag-ingestion-pipeline).
• If user gives domain-specific terminology or language preferences: add addon-domain-semantic-adaptation.

Observability Intensity Examples

• Baseline runtime default: Build a public FastAPI service for customer accounts with auth and Postgres. Route to the Python base stack and include the runtime-specific observability add-on only.
• Explicit full telemetry push: Build a public FastAPI service and add structured logs, metrics, tracing, health checks, and redaction rules. Route to the Python base stack and include addon-observability-telemetry plus the runtime-specific observability add-on.
• Cross-runtime observability policy: Define a shared observability policy we can apply across our Python API and Next.js app before we choose implementation details. Add addon-observability-telemetry first, then add runtime-specific observability add-ons only after a concrete runtime is selected.

Default Compositions

For public api and web flows in production-default mode, append the runtime-specific observability add-on as baseline observability unless the user explicitly opts out.

• Python PDF/Markdown RAG worker: architect-python-uv-batch + addon-rag-ingestion-pipeline + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Python legal PDF clause RAG worker: architect-python-uv-batch + addon-rag-ingestion-pipeline + addon-docling-legal-chunk-embed + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Python API with RAG endpoints: architect-python-uv-fastapi-sqlalchemy + addon-rag-ingestion-pipeline + addon-observability-python-structlog + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• FastAPI with auth and access control: architect-python-uv-fastapi-sqlalchemy + addon-auth-access-control + addon-observability-python-structlog + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Python observability-first flow: architect-python-uv-fastapi-sqlalchemy + addon-observability-telemetry + addon-observability-python-structlog + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js observability-first flow: architect-nextjs-bun-app + addon-observability-telemetry + addon-observability-nextjs-pino + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js on Vercel baseline: architect-nextjs-vercel-app + addon-observability-nextjs-pino + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js Nostr client: architect-nextjs-bun-app + architect-nostr-intent-router + addon-nostr-client-nextjs + addon-nostr-nip-profile-selector + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js Nostr client with social feed UI: architect-nextjs-bun-app + architect-nostr-intent-router + addon-nostr-client-nextjs + addon-nostr-nip-profile-selector + ui-social-feed-shell + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js typography-first journal with NIP-23 + Ancient Greek translation: architect-nextjs-bun-app + architect-nostr-intent-router + addon-nostr-client-nextjs + addon-nostr-nip-profile-selector + addon-nostr-key-custody + addon-nostr-nip23-longform + addon-llm-ancient-greek-translation + ui-editorial-writing-surface + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js document intake UI for RAG: architect-nextjs-bun-app + addon-rag-ingestion-pipeline + ui-document-intake-dropzone + addon-observability-nextjs-pino + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Full-stack Next + Prisma + vector search: architect-next-prisma-bun-vector + addon-observability-nextjs-pino + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate (+ addon-rag-ingestion-pipeline if ingest/retrieval pipeline needed)
• FastAPI with direct LangChain LLM endpoints: architect-python-uv-fastapi-sqlalchemy + addon-langchain-llm + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• FastAPI with LangGraph agent orchestration: architect-python-uv-fastapi-sqlalchemy + addon-langgraph-agent + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Cross-architecture translation flow: architect-nextjs-bun-app or architect-python-uv-fastapi-sqlalchemy + addon-llm-translation + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Cross-architecture direct LLM SDK flow: architect-nextjs-bun-app or architect-python-uv-fastapi-sqlalchemy + addon-direct-llm-sdk + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Cross-architecture LangChain LLM flow: architect-nextjs-bun-app or architect-python-uv-fastapi-sqlalchemy + addon-langchain-llm + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Cross-architecture LangGraph agent flow: architect-nextjs-bun-app or architect-python-uv-fastapi-sqlalchemy + addon-langgraph-agent + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Google ADK agent flow: architect-nextjs-bun-app or architect-python-uv-fastapi-sqlalchemy + addon-google-agent-dev-kit + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate
• Next.js Vercel AI Labs flow: architect-nextjs-bun-app + addon-vercel-ai-labs + addon-decision-justification-ledger + addon-domain-semantic-adaptation + addon-deterministic-eval-suite + addon-human-pr-review-gate

Output Template

Use this response shape:

Selected skills (in order):
1) <skill-name> - <reason>
2) <skill-name> - <reason>

Install:
npx skills add ajrlewis/ai-skills --skill <skill-name>
npx skills add ajrlewis/ai-skills --skill <skill-name>

Execution order:
1) Apply base architect skill.
2) Apply add-ons.
3) Run each skill's validation checklist.
4) Summarize generated artifacts and any blockers.

Execution mode:
- full | offline-smoke | limited-by-sandbox

Containerization:
- required across runnable base architect skills (default) | explicit-no-docker (user requested)

Human review gate:
- required (default) | explicit-opt-out (user requested)

Decision justification ledger:
- required (default) | non-optional policy

Semantic adaptation:
- required (default) | explicit-opt-out (user requested)

Observability telemetry:
- required-for-public-api-web (default) | explicit-opt-out (user requested)

Observability intensity:
- baseline-runtime-default (default) | explicit-full-telemetry-push (user requested or clearly primary requirement)

Deterministic eval suite:
- required (default) | explicit-opt-out (user requested)

LLM judge evals:
- optional (default) | explicit-required (user requested)

Guardrails

• Documentation contract for generated code:

- Python: write module docstrings and docstrings for public classes, methods, and functions. - Next.js/TypeScript: write JSDoc for exported components, hooks, utilities, and route handlers. - Add concise rationale comments only for non-obvious logic, invariants, or safety constraints. - Apply this contract even when using template snippets below; expand templates as needed.

• Prefer the smallest skill set that satisfies requirements.
• Do not select multiple base architects unless user explicitly requests a multi-repo architecture.
• architect-nostr-intent-router is a protocol control-plane architect and may be combined with one base architect.
• If a requirement conflicts with a selected skill, call it out and suggest the nearest valid composition.
• Surface blockers early (missing uv/bun, no network, restricted Docker) and include an actionable fallback.
• Never silently downgrade from production-default to no-docker behavior.
• Do not invent prompt-specific one-off skills during normal routing; compose generic architect-*, addon-*, and reusable ui-* skills.
• Prefer generic UI skills (ui-social-feed-shell, ui-editorial-writing-surface) with style inputs over brand-locked UI skill creation.
• Default to addon-human-pr-review-gate for production scaffolds.
• Default to addon-decision-justification-ledger for production scaffolds.
• Default to addon-domain-semantic-adaptation so user/domain language is reflected in output copy and docs.
• Default to addon-deterministic-eval-suite for hard pass/fail gating.
• Treat addon-llm-judge-evals as secondary quality scoring unless user explicitly makes it blocking.
• No skill decision should be returned without a concrete justification and trace entry.

Decision Justification Rule

• Every non-trivial decision must include a concrete justification.
• Capture the alternatives considered and why they were rejected.
• State tradeoffs and residual risks for the chosen option.
• If justification is missing, treat the task as incomplete and surface it as a blocker.