Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/tomevault-io/skills-registry --skill agent-workflow명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
| Use when this capability is needed.
> Use when this capability is needed.
Review architecture and API design for the vfs-s3 project. Use when the user mentions @architect, asks to review an issue's design, discuss module boundaries, API shape, or architectural decisions for vfs-s3. Also trigger when the user wants to create an ADR (Architecture Decision Record) or evaluate a technical approach for the project. Intended for dispatch from Codex automation or Claude routines; GitHub trigger phrase: @vfs-s3-bot please prepare design doc Use when this capability is needed.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | agent-workflow |
| description | >- Use when this capability is needed. |
Truth: package/INVOCATION.md after install. Reference: reference.md. Manual: package/AGENTWORKFLOW_MANUAL.html for humans; do not load it into model context unless the user asks.
This skill is a workflow gate, not a documentation bundle. Keep context small:
SKILL.md at skill activation.package/INVOCATION.md only when a step is ambiguous or a command is needed.reference.md only for command lookup.aw status, aw next, aw task brief, aw context plan, aw sync inbox) over opening full ledgers.When the user says 更新 aw, update aw, 升级 aw, or asks to refresh an already installed skill, run the project-local updater:
./scripts/aw upgrade --from-github --adapters
./scripts/aw update --from-github --adapters is the same path. This fetches https://github.com/duntak1/agentworkflow.git, removes the old local skill install, installs the fresh skill, and replaces the current project's agent-workflow/ package and scripts/ while preserving business docs/, reference/, and workflow state. Use --repo <url> or --ref <branch-or-tag> only when the engineer explicitly asks for a fork, branch, or release tag.
When the user says 启动 aw, @aw, aw start, or equivalent:
1=产品2=前端3=后端4=全栈aw project scan, summarize docs/PROJECT_SCAN.md, and ask the engineer to confirm new vs existing.aw config init --sync-center 1|2|3.全栈 means frontend/backend code is in one repository by default, build target 3=fullstack; sync center is optional unless the engineer confirms split repos, different computers, or PM-managed collaboration.
Default budgets per AT-T:
| Context | Budget |
|---|---|
| Business files | 6 files |
| Symbols | 12 symbols |
| Precise searches | 3 searches |
| Handoff read | latest summary only, not whole history |
| Sync inbox read | relevant peer/event only, not all snapshots |
| Reference docs | manifest + explicitly cited files only |
Do not exceed a budget without saying why and asking the engineer to confirm the expansion. Never do an aimless full-repo scan. Avoid generated/cache/dependency/build/log directories: .git, node_modules, dist, build, coverage, .next, .nuxt, target, vendor, tmp, logs.
Use this lean order:
./scripts/aw status --json
./scripts/aw next
./scripts/aw task brief <AT-T>
./scripts/aw context plan --task <AT-T>
./scripts/aw context gate --task <AT-T>
Read only files listed in docs/context/tasks/CTX-<AT-T>.md. Prefer CodeGraph / aw code-map / aw context symbol, caller/callee, impact, and affected-test queries. If CodeGraph is unavailable, use CODE_MAP, CODE_CONTEXT_INDEX, FILE_INDEX, and precise rg.
Code map is automatic by default. aw context plan, aw task start, aw task complete, aw watch index, and aw gate pre-commit refresh/check docs/context/CODE_MAP.md unless AW_CODE_MAP_AUTO=0 is set for an explicit exception. Before coding in existing or large projects, query the map instead of scanning the repo:
./scripts/aw code-map build
./scripts/aw code-map query "<feature|symbol|route>"
./scripts/aw code-map impact "<symbol|route>"
CODE_MAP.md is a locator index, not permission to read full files. Convert findings into the task Context Plan and wait for confirmation before opening file contents.
| Stage | Hard rule |
|---|---|
| DSL | DSL status not 已审 → no business code |
| Plan | Plan status not 可执行 → no business code |
| Confirm | Run aw confirm <dsl> <plan> before AT-T execution |
| Task start | aw next → aw task brief <id> → ask scope / acceptance / non-goals → aw task confirm <id> "已确认:范围=...;验收=...;非目标=..." |
| Context | aw context plan --task <id> → engineer reviews allowed files → aw context gate --task <id> |
| Coding | Only after aw task start <id> may you request/use aw paste task or edit business code |
| Completion | aw task complete <id> verifies; failed verification logs Bug and keeps task open |
| Checkpoint | Completed AT-T needs Git decision, compact/handoff evidence, and docs/FILE_INDEX.md refresh when business files changed |
When the user says “执行研发任务”, “开始开发”, “做下一个任务”, or equivalent, do not start coding. Start with aw next and aw task brief.
| Lane | Use for | Do not use for |
|---|---|---|
Handoff (docs/handoff/PROJECT_HANDOFF.md) | Current goal, progress, blockers, next 1-3 steps | Long-term reusable facts |
Memory (docs/memory/) | Stable decisions, preferences, reusable procedures, recurring risks, summarized chat episodes | Raw chat logs or transient task status |
Use aw compact "focus" --write --snapshot before Codex context compaction, new chats, model switches, long pauses, or after a large requirement/AT-T batch. Add --memory-summary only for reusable decisions or follow-ups. Do not store secrets.
New session lean resume:
./scripts/aw handoff --check
./scripts/aw memory inject
./scripts/aw status --json
./scripts/aw next
Read handoff/memory selectively. Do not read ENGINEERING_INDEX.md into AI context; it is for human engineers.
aw dsl suite <slug> "title" creates multi-file DSL coverage for requirements, pages, interactions, events, boundaries, and acceptance.aw dsl review <dsl> --write creates the engineer review package before aw approve dsl.aw approve dsl <dsl> --plan may guide Plan generation; PM/three-side projects should use aw pm plan --write before dispatch.aw plan change, aw plan task-add, and aw task split handle development-time changes without bloating the active task.aw pm start and aw pm init <project-harness>; PM owns shared references, Pencil intake, shared DSL/Plan, dispatch, dashboard, and lifecycle gates.For split frontend/backend repositories:
aw sync init <project-harness> --project ... --agent ... --role ....aw sync pull + aw sync gate --task <id>.aw sync inbox --from <peer> summaries and relevant event/contract files.aw sync pull imports read-only inbox files; never treat inbox as local truth until adopted through REQ / Bug / Plan / Handoff.aw contract change, aw contract test, and aw contract gate.aw agents register; common roles can be bootstrapped with aw agents register --defaults.aw agents claim, aw agents heartbeat, and aw agents gate --strict.| Principle | Meaning |
|---|---|
| Think before coding | Clarify authority, scope, acceptance, and non-goals |
| Simplicity first | Keep changes small and directly tied to the AT-T |
| Mature solutions first | Prefer proven libraries, SDKs, framework features, and official patterns |
| Surgical changes | Avoid unrelated refactors and noisy diffs |
| Goal-driven execution | Verify against real project commands and recorded acceptance |
Also maintain completeness, traceability, maintainability, and handoffability. Use aw audit, aw policy, aw security, aw service-catalog, aw release, aw trace check, aw metrics summary, aw ops gate, aw report handoff, aw report release, aw score record, and aw recover ... when the task touches risk, production, dependencies, services, release, multi-agent work, or handoff.
aw req new; record development-time changes with aw req change, then update DSL/Plan/ATOMIC before continuing.aw bug add.docs/FILE_INDEX.md when adding/deleting/renaming business files.aw commit --task <id> --changelog "..."; do not commit automatically.Pencil design files are formal inputs. Store .pen sources under global/references/design/pencil/source/, exports under exports/, screenshots under screenshots/, and changes in DESIGN_CHANGELOG.md. Do not parse encrypted .pen files as plain text; use Pencil tooling or exported artifacts.
Source: duntak1/agentworkflow — distributed by TomeVault.