Skip to main content

using-avid

Use when starting product, feature, build, stack, research, agent-framework, or implementation work. Loads AVID personal workflow defaults and skill routing. Defaults are advisory and scene-dependent — not absolute hard blocks. Skip or adapt when the deployment target or user intent makes a default a bad fit.

소스 정보

저장소
AVIDS2/avid-skill
최근 소스 활동
2026년 7월 24일 01:13
감지된 SKILL.md 언어
영어
스타
1
포크
0

설치 방법

기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.

소스 파일 검토

설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.

파일 탐색기
3 개 파일

SKILL.md 표시 중

SKILL.md
소스 지침 · 읽기 전용 미리보기
name
using-avid
description
Use when starting product, feature, build, stack, research, agent-framework, or implementation work. Loads AVID personal workflow defaults and skill routing. Defaults are advisory and scene-dependent — not absolute hard blocks. Skip or adapt when the deployment target or user intent makes a default a bad fit.
# Using AVID You have the **avid-skill** plugin: a **personal product-engineering workflow**, not a replacement for official tools (Tavily, Context7, Playwright, Matt Pocock skills, Impeccable, taste-skill, etc.). AVID exists to reduce blind builds and wrong defaults. It does **not** exist to force one stack/runtime into every project. ## Priority 1. **User's explicit instructions** — always highest 2. **Scene fit / technical reality** — invalid defaults must be dropped (e.g. Pi on pure web FE/BE) 3. **AVID suggested defaults** — prefer when they fit 4. **Claude Code process:** Matt Pocock skills (not Superpowers) + **ponytail** when coding 5. Default system behavior ## How to apply (important) - Treat the table below as **recommended**, not court orders. - When a skill fits: load it and follow the useful parts. - When it doesn't: **skip or adapt**, and say one line why (scene mismatch / user override / cost). - Never "comply" into a catastrophic wrong architecture just to satisfy a checklist. - Tiny fixes, refactors, and already-scoped product work: no need to run bootstrap theater. ## Suggested routing | Situation | Suggested skill | Notes | |-----------|-----------------|-------| | New product / new vertical / greenfield | `product-bootstrap` | Strongly recommended; still skippable if user already has research/spec | | About to self-build a capability | `research-before-build` | Prefer find-before-found | | Missing installable tool | `install-dont-reinvent` | Prefer install; hand-roll only with reason | | Stack choice | `stack-defaults` | Defaults with easy exceptions | | Agent runtime | `agent-framework-choice` | **Scene-based** — see below | | Knowledge / RAG | `knowledge-architecture` | Clarify consumers first | | High-stakes decision | `multi-agent-deliberation` | Optional second opinions | ## Agent runtime — scene fit (do not hard-force Pi) | Scene | Prefer | Avoid forcing | |-------|--------|----------------| | Web FE/BE product, API, SaaS UI | Thin LLM SDK / server agent framework as needed; often **LangGraph** only if workflow/HITL graph is real | **Pi SDK as default** — wrong shape for normal web apps | | Local coding / desktop / Electron / workspace harness agent | **Pi** family: Pi SDK, pi-ai, or embed/use Pi agent runtime — pick what docs + product need | LangGraph cargo-cult for a simple tool loop | | Multi-step business workflow, durable state, HITL approvals | **LangGraph** + langgraph skills/docs | Pi-as-graph | | "Just chat + a few tools" | Official provider SDK loop first | Heavy frameworks | Always: fresh docs (Context7 / official) before coding on a chosen framework. **Model API protocol (when building agents):** prefer **OpenAI Chat Completions** (`/v1/chat/completions`) for multi-model / gateway / Pi `openai-completions` compatibility. Treat Responses / Anthropic Messages as optional vendor paths — not the only P0 protocol unless the product is single-vendor by design. See `agent-framework-choice`. ## Tool defaults (prefer official tools) | Need | Prefer | |------|--------| | Web research | Tavily | | Vertical search | AnySearch | | Lib docs | Context7 | | Browser QA | Playwright | | Email / pay / data | Resend / Stripe / SQLite→Postgres(Supabase) **when they fit** | | Agent LLM wire protocol | **Chat Completions** first (multi-provider) | | Minimal code | ponytail | | Process (Claude Code) | **Matt Pocock skills** (`grill-me` / `grill-with-docs` → `to-spec`/`to-tickets` → `implement`/`tdd`/`code-review`) — **not Superpowers** | | Frontend UI | **Impeccable + taste-skill** (`design-taste-frontend`) together — taste for direction/anti-slop; Impeccable for craft/audit/polish | ## Frontend pair (advisory) - Non-trivial UI: load **both** `design-taste-frontend` (taste) and **Impeccable** (`/impeccable …`), not only one. - Optional direction skills (`high-end-visual-design`, etc.) when the aesthetic is already chosen. - shadcn / component libs remain implementation tools under that pair. ## Anti-patterns - Blind build with zero market scan on a **new product** when research is cheap - Forcing Pi SDK because "AVID says harness → Pi" on a web platform where it doesn't apply - Forcing LangGraph on a trivial tool loop - Silent hand-roll of a mature hosted service without a reason - Ritual checklists that block a clearly scoped hotfix - Using **Superpowers** as the CC engineering methodology (superseded by Matt Pocock skills) - Shipping UI with only Impeccable **or** only taste when both are available ## After planning Prefer minimal implementation (ponytail); on Claude Code prefer Matt Pocock process skills when they fit. Bias to P0. User can always say "skip AVID for this" — obey.
GitHub에서 보기