AGENT SKILL

efficient

Use to identify the fastest safe reuse-first approach (e.g. extend existing Supabase RPC instead of a new service). Not for a deep source-receipt audit (ponytail) or skipping necessary security tests.

License 许可
MIT
所属仓库
amoai-tech/mdeai
最近更新
2026-11-13
Stars
★ 2
Forks
1

01 前言 Front Matter

name
efficient
description
Use to identify the fastest safe reuse-first approach (e.g. extend existing Supabase RPC instead of a new service). Not for a deep source-receipt audit (ponytail) or skipping necessary security tests.

03 SKILL.md 正文

EFFICIENT — Smallest Safe Path

Goal Complete the task with the least unnecessary work while preserving correctness, security, and maintainability.

Rules

  • Always ask: Is there a better, faster, simpler, safer, or more efficient way to complete this?
  • If yes, use it.
  • Verify current project/runtime truth before designing new work.
  • Prefer reuse over new code.
  • Do not duplicate services, schemas, state, utilities, agents, workflows, or infrastructure.
  • Do not inspect, research, test, or modify unrelated areas.
  • Do not over-engineer for hypothetical future needs.
  • Do not trade correctness or security for speed.

Subtract before adding (P Stack): First ask whether existing configuration, a single canonical owner, or removal of redundant logic solves the problem. Compare the smaller alternative against the requested outcome; don't build a parallel abstraction or orchestrate multiple agents merely to appear thorough. Use additional reviewers only when risk or evidence justifies their cost.

Use / do not use: Use for “Find a faster way to deliver this MDE fix using existing services.” Do not use for “Trace exact library/source receipts” (ponytail) or as a reason to omit technical authorization or necessary tests.

Canonical-owner check: Before proposing any new module, search only the relevant route/RPC/component/agent and its callers; name the existing owner and why it cannot be reused if custom code is unavoidable. Do not explore unrelated directories or create a second engine.

Preferred order

existing implementation → configuration → installed/vendor capability → official SDK/module → official example/template → minimal custom code

Use custom code only for the remaining gap.

Research Research only when external/current behavior affects the decision. Prefer:

  1. installed source/types
  2. official docs
  3. official repository/examples
  4. maintained first-party references

For every important external reference used:

URL → COPY / ADAPT / MODEL / REFERENCE ONLY → exact destination → verification

Do not create link dumps.

Implementation

  • Make the smallest focused change.
  • Preserve existing contracts and sources of truth.
  • Reuse existing tests/components/functions where possible.
  • Delete superseded or duplicated logic when safe.
  • Avoid unrelated refactors.

Verification Use the cheapest reliable proof first:

focused test → integration/security test → typecheck/lint → build → browser/E2E → production proof when required

Stop when the required outcome is proven. Do not run broad expensive checks without a reason.

Stop when: The smallest existing implementation or supported extension is identified and its risk-matched proof is known. Execute the authorized change or hand the concrete plan to implement; do not repeat an architecture audit after the decision is supported.

Decision Before finishing, state:

  • simpler/faster approach used
  • custom work avoided
  • remaining risk or blocker
  • single highest-value next step

Do not claim 100% correct or production-ready without evidence.

04 同仓库更多 Skills