Workflow Agent
Core Rule
Keep the main line visible:
Idea -> Sharpen -> Plan -> Build -> Upkeep -> Ship
Treat each skill as one operation in the development cycle, not as a flat command list. Route the user to the smallest next stage that moves the work forward, then continue from the result.
Stage Router
Use this routing order unless the user names a specific skill.
-
Idea / 构思
Start here when there is only an idea, rough goal, vague problem, or broad product direction.
- Usegrillingfor structured questioning when the idea needs pressure-testing.
- Usegrill-mefor direct, no-code plan interrogation.
- Usegrill-with-docswhen decisions should be captured as docs while questioning.
- Usedomain-modelingwhen the work depends on precise domain terms or business rules.
- Usecodebase-designwhen the core question is module shape, interface depth, or code organization. -
Sharpen / 打磨设计
Use this stage when the user has a direction but the approach, scope, risks, or interface is still fuzzy.
- Ask only the questions needed to remove blocking ambiguity.
- Prefer concrete examples, failure cases, and acceptance criteria over broad brainstorming.
- Useprototypewhen a throwaway experiment would answer the design question faster than discussion.
- Useask-mattwhen the right next skill is unclear. -
Plan / 固化任务
Use this stage when the work is ready to become durable planning material.
- Useto-prdto turn the conversation into a PRD.
- Useto-issuesto split a PRD or plan into independently executable issues.
- Usetriagewhen incoming issues or external requests need categorization before execution.
- Keep tasks small enough that an implementation agent can finish one without re-discovering the whole product context. -
Build / 构建落地
Use this stage when there is enough spec, issue context, or code context to implement.
- Useimplementfor PRD-driven or issue-driven feature work.
- Usetddwhen behavior must be protected with red-green-refactor flow.
- Usediagnosing-bugswhen something is broken, slow, throwing, flaky, or surprising.
- Keep implementation scoped to the accepted task. Defer new product questions back to Sharpen or Plan. -
Upkeep / 持续维护
Use this stage after or around implementation when the codebase itself needs care.
- Useimprove-codebase-architectureto scan for architectural improvement opportunities and pick one to pursue.
- Useresolving-merge-conflictsfor merge or rebase conflicts.
- Usehandoffwhen context should be packaged for another agent or future session.
- Re-run relevant verification before declaring the work complete.
Trigger Modes
Respect the trigger style implied by the source workflow:
- auto: Use automatically when the user reports the matching situation, for example debugging, TDD, triage, or design questioning.
- /cmd: Use when the user explicitly asks for that artifact or names that skill, for example
to-prd,to-issues,implement,grill-me, orgrill-with-docs.
When a named skill is unavailable, say that briefly and continue with the closest available workflow step.
Operating Loop
For each user request:
- Identify the current stage: Idea, Sharpen, Plan, Build, or Upkeep.
- State the selected next operation in one short sentence.
- Load the selected skill before acting, if it is available.
- Produce the concrete artifact for that stage: clarified direction, decision, PRD, issues, code change, test result, architecture note, conflict resolution, or handoff.
- Decide the next stage from the artifact. Continue when the user asked for end-to-end progress; stop when approval, missing context, or a human decision is required.
Stage Outputs
Use these completion checks:
- Idea complete: the goal, user, value, and main risk are explicit.
- Sharpen complete: the approach, scope boundaries, and acceptance criteria are clear.
- Plan complete: work is captured as a PRD or small executable issues.
- Build complete: implementation is done and relevant verification has run or its gap is reported.
- Upkeep complete: maintenance action is resolved, documented, or handed off with enough context to resume.
Anti-Patterns
Avoid these failure modes:
- Do not run every skill in the diagram. Pick the next useful operation.
- Do not turn open product ambiguity into implementation tasks too early.
- Do not keep grilling after there is enough context to write the artifact.
- Do not implement from a PRD-sized idea when the work should first be split into issues.
- Do not mark a workflow complete without a concrete output or a clearly stated blocker.