architect-conventions

Shared conventions for every SDLC development-phase architect agent: hard rules, code quality bar, workflow steps (superpowers invocation, spec reading, codebase exploration, verification), and the re

SKILL.md

GitHub

Architect Conventions (shared)

Every development-phase architect follows these rules. Your own agent definition adds stack-specific detection, conventions, and verification commands on top — where the two conflict, your agent definition wins.

Hard rules

  • Never delete files unless the spec explicitly asks for it.
  • Never modify .env, secrets/*, or ~/.claude/**.
  • Never disable existing tests to "make them pass". Mark as skip with a code comment if you genuinely can't fix in scope, and report it in your summary.
  • Never push branches or open PRs — that's the documentation phase's job.
  • Never add a dependency not declared in the BA spec or required by your implementation. Justify additions in DECISIONS.
  • Never edit lockfiles (package-lock.json, yarn.lock, pnpm-lock.yaml, composer.lock, poetry.lock, uv.lock, etc.) by hand.

Code quality bar

  • Follow existing patterns. Don't introduce a new way of doing things in scope of this feature.
  • No TODO/FIXME comments unless explicitly noting future work agreed upon by BA.
  • No commented-out code blocks.
  • No "in case we need it later" abstractions. YAGNI.
  • New deps via the project's package manager, pinned to a sane range (^x.y.z or stack equivalent). Never * or latest.
  • Match the existing styling/formatting approach. Don't introduce a new one.

Workflow steps

The orchestrator dispatches you in one of two passes: planning or implementation. Its base prompt tells you which pass you're in. In both:

  1. If superpowers is installed (no superpowers_unavailable flag in CONTEXT), invoke superpowers:using-superpowers via the Skill tool to discover available skills.
  2. Read the spec at docs/plans/{task_slug}/01-business-analysis.md (plus any earlier-aspect outputs listed in your CONTEXT trailer).
  3. Detect project shape — start from the project_shape: block in your CONTEXT trailer (package manager, runner prefix, test / lint / typecheck / format commands, test framework, CI, monorepo), computed once per session by scripts/detect-stack.sh. Run your agent definition's detection checklist only for keys that are null there — re-deriving a known answer is paid in tokens on every pass.
  4. Explore the codebase — Glob/Grep for the most similar existing feature; Read actual files and mirror their naming and patterns.
  5. Read CLAUDE.md — project conventions are sacred.
  6. Implement. Use Edit for existing files, Write for new ones. Keep changes minimal.
  7. Invoke convention skills the orchestrator passed that are relevant to the task.
  8. Verify with your agent definition's stack-specific commands (build, compile, import smoke test). Failures block completion. Formatters and the batched typecheck/lint are hook-driven — the stack plugin's PostToolUse hook formats every file you write and its SubagentStop hook runs the typecheck or linter once after you return, relaying failures to the orchestrator as a retry hint — so do not run those yourself.
  9. If superpowers is installed, invoke superpowers:verification-before-completion via the Skill tool before returning.

Deliverable contract

Write the detailed implementation report to docs/plans/{task_slug}/02-development{-aspect_suffix}.md covering: files created/modified (with purpose), dependencies added (with why), detected project shape, key design decisions with rationale, deviations from spec, manual verification performed, and open issues/blockers for the next phases.

Return value (COMPACT summary)

Return ONLY (≤3K tokens):

FILES CREATED: [paths with type tag]
FILES MODIFIED: [paths]
DEPS ADDED: [package@version, ... or "none"]
PROJECT SHAPE: [one line of detected key=value pairs]
DECISIONS: [3-5 bullets]
BLOCKERS: [empty or up to 3 lines]