Interface-first delegation

Principles

• Humans own strategic design quality; agents handle tactical implementation when boundaries and feedback are clear.
• Invest in design every day—do not treat code as disposable black noise.
• Plans should name module/interface deltas precisely; avoid narrative filler (token-efficiency output awareness).

Workflow

1. For non-trivial work, the plan or PRD must list:
   • Which modules change
   • How public interfaces evolve (add/change/remove)
   • What tests prove the interface contract
2. Treat deep modules as gray boxes when risk allows: own the interface and tests; delegate the interior.
3. When delegating to agents or subagents, give each task a spec plus an executed check that decides pass/fail—never accept the agent's own report (verified-delegation).
4. Never hand-wave high-risk domains (auth, money, data loss, security, migrations)—keep human review explicit.
5. Keep a module map visible during planning and implementation.
6. Put module and interface terms in the glossary (ubiquitous-language).

Anti-pattern

Specs-to-code that divests from system design: iterating only on documents while the codebase decays.

Rules

• Prefer the smallest useful artifact: an interface delta and failing contract test, not a sprawling plan.
• Implementation may proceed only after interface and verification intent are clear.
• Chain with tdd-feedback-loops so delegated interiors stay verified.

Sibling skills

• Align on the design concept with grill-me before naming interfaces.
• Prefer deep boundaries from deep-modules-architecture as the units of delegation.
• Shared vocabulary from ubiquitous-language keeps plans and code aligned.
• Scale delegation to multiple workers with verified-delegation: spec + executed check per task.