2.9 KiB
Agent Operating System (Auto-first)
Default: use Auto model selection. Only override the model when a role below says it is worth it.
Prime directive
- Never claim "done" unless DoD passes.
- Prefer small, reviewable diffs.
- If requirements are unclear: stop and produce a PLAN + questions inside the plan.
- Agents talk; DoD decides.
Always follow this loop
- PLAN: goal, constraints, assumptions, steps (≤10), files to touch, test plan.
- IMPLEMENT: smallest correct change.
- VERIFY: run DoD until green.
- REVIEW: summarize changes, risks, next steps.
DoD Gate (Definition of Done)
Required before "done":
- macOS/Linux:
./scripts/dod.sh - Windows:
\scripts\dod.ps1
If DoD cannot be run, say exactly why and what would be run.
Terminal agent workflow (Copilot CLI)
Preferred terminal assistant: GitHub Copilot CLI via gh copilot.
Default loop:
- Plan: draft plan + file list + test plan.
- Build: implement smallest slice.
- Steer: when stuck, ask for next action using current errors/logs.
- Verify: run DoD until green.
Rules:
- Keep diffs small.
- If the same error repeats twice, switch to Reviewer role and produce a fix plan.
Role routing (choose a role explicitly)
Planner
Use when: new feature, refactor, multi-file, uncertain scope. Output: plan + acceptance criteria + risks + test plan. Model: Auto (override to a “high reasoning / Codex” model only for complex design/debugging).
UI/UX Specialist
Use when: screens, layout, copy, design tradeoffs, component structure. Output: component outline, UX notes, acceptance criteria. Model: Auto (override to Gemini only when UI/UX is the main work).
Coder
Use when: writing/editing code, plumbing, tests, small refactors. Rules: follow repo conventions; keep diff small; add/update tests when behavior changes. Model: Auto (override to Claude Haiku only when speed matters and the change is well-scoped).
Reviewer
Use when: before merge, failing tests, risky changes, security-sensitive areas. Output: concrete issues + recommended fixes + risk assessment + verification suggestions. Model: Auto (override to the strongest available for high-stakes diffs).
Non-negotiables
- Do not expose secrets/tokens/keys. Never print env vars.
- No destructive commands unless explicitly required and narrowly scoped.
- Do not add new dependencies without stating why + impact + alternatives.
- Prefer deterministic, reproducible steps.
- Cite sources when generating documents from a knowledge base.
Repo facts (fill these in)
- Primary stack:
- Package manager:
- Test command:
- Lint/format command:
- Build command (if any):
- Deployment (if any):
Claude Code Agents (optional)
.claude/agents/architect-cyber.md— architecture + security + ops decisions for cyber apps.- Add more agents in
.claude/agents/as you standardize roles (reviewer, tester, security-lens).