Writing a rolling-wave phased implementation plan from a spec before coding: a phase-epic-task hierarchy where Phase 1 is detailed into dispatch-ready tasks and later phases stay epic-level for elaboration during execution. Use when a multi-file feature needs decomposition; runs after ring:exploring-codebases or pre-dev gates, hands off to ring:executing-plans or ring:running-dev-cycle. Skip for single-file changes or spikes.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Writing a rolling-wave phased implementation plan from a spec before coding: a phase-epic-task hierarchy where Phase 1 is detailed into dispatch-ready tasks and later phases stay epic-level for elaboration during execution. Use when a multi-file feature needs decomposition; runs after ring:exploring-codebases or pre-dev gates, hands off to ring:executing-plans or ring:running-dev-cycle. Skip for single-file changes or spikes.
Writing Plans
When to use
Spec or requirements exist for a multi-step task and no implementation has started
Feature spans multiple files/layers and needs decomposition before coding
Handing off implementation to a separate session, agent, or human
Skip when
Single-file change with obvious shape (just do it)
Exploratory spike — phased plans assume known requirements
Spec is still in brainstorming; the plan would lock premature decisions
Sequence
Runs after: ring:exploring-codebases, ring:planning-large-features (gates 0-6 artifacts) or ring:planning-small-features (gates 0-2 artifacts) — their outputs feed the spec
Runs before: ring:executing-plans (rolling-wave execution) or ring:running-dev-cycle (gated subagent workflow)
Write the plan assuming the implementer is skilled but has zero context for this codebase, toolset, or problem domain.
The plan is a rolling-wave document. Only the first phase is detailed to task level at plan time; later phases stay at epic level until execution reaches them. Detail decays: code written in Phase 1 invalidates assumptions baked into Phase 3 tasks, so do not write Phase 3 tasks yet. ring:executing-plans elaborates each subsequent phase against the codebase as it actually exists.
Announce at start: "Using ring:writing-plans to author the implementation plan."
Default save path:docs/plans/YYYY-MM-DD-<feature-name>.md
(User preferences override.)
Invoked from pre-dev: when dispatched as the final gate of ring:planning-large-features or ring:planning-small-features, the spec inputs are the pre-dev artifacts — trd.md (plus feature-map.md, openapi.yaml, the schema file, and dependencies.md on the Large track). On Large, plan phases MUST mirror feature-map.md phases one-to-one. Output path is docs/pre-dev/{feature}/plan.md, overriding the default above; standalone invocations keep the default. plan.md is always a SINGLE document per feature: on multi-module topologies (monorepo fullstack / multi-repo), each epic carries one line **Target:** backend | frontend | infra (placed right before **Status:**); for multi-repo features the orchestrator copies plan.md into each repo and the local dev-cycle executes only epics whose Target matches that repo. No per-module plan splits.
Plan Language
Before authoring, ask which language the plan's prose should be written in — using the question/ask tool the harness provides (e.g. AskUserQuestion in Claude Code). Offer three options: English (default), Brazilian Portuguese (pt-BR), Spanish. If the user skips or no question tool is available, default to English.
The choice covers narrative prose only — Goal, Architecture, Context, Implementation vision, and other descriptions. Everything a downstream skill or agent parses stays verbatim English regardless of choice: section headers and format keywords, **Status:** values (Pending/Doing/Done/Failed), Phase-Overview status cells (Detailed/Epic-level/Complete), **Target:** values, file paths, commands, code snippets, and identifiers. Translating those breaks ring:executing-plans and ring:running-dev-cycle status matching.
Standards
Do NOT fetch standards documents while planning — standards compliance is enforced by the implementation agents and reviewers downstream. Plans reference DRY, YAGNI, and TDD generically.
Blocker — STOP and Report
Do not write a plan on a shaky foundation. STOP and ask when:
Situation
Action
Vague requirements ("make it better", "add feature")
STOP. Ask: "What specific behavior should change?"
Missing success criteria
STOP. Ask: "How do we verify this works?"
Unknown codebase structure (can't locate files)
STOP. Run ring:exploring-codebases first, then plan
Conflicting constraints
STOP. Ask: "Which constraint takes priority?"
Multiple valid architectures without guidance
STOP. Ask: "Which pattern should we use?"
Scope Check
If the spec covers multiple independent subsystems, suggest breaking it into separate plans — one per subsystem. Each plan must produce working, testable software on its own.
If brainstorming already split the spec into sub-project specs, write one plan per sub-spec.
Plan Hierarchy
Level
Granularity
When detailed
Phase
Independently verifiable milestone — software works at the end of every phase
At plan time
Epic
Cohesive unit of work inside a phase (one capability, one subsystem)
Phase 1 at plan time; later phases during execution (rolling wave)
Rules:
Every phase ends with working, testable software. No phase ends mid-refactor.
2–5 epics per phase. An epic that needs more than a paragraph to describe is two epics.
Order phases by dependency first, then by risk — front-load whatever invalidates the design if it turns out wrong.
Code Snippet Policy
Default is prose, not code. Describe intent, decisions, and shape; the implementer writes the code at execution time with the real codebase in front of them.
Include a snippet ONLY when prose cannot pin down the decision:
Justified
Example
Public contract other epics depend on
API signature, event schema, migration DDL
Non-obvious algorithm where the approach IS the decision
Custom balancing logic, conflict-resolution rule
Exact artifact where approximation breaks behavior
Config block, regex, SQL query
If the snippet exists to "save the implementer time", delete it. If it exists because two epics would otherwise disagree about a contract, keep it.
Plan Document Header
Every plan MUST start with this header:
# [Feature Name] Implementation Plan> **For implementers:** Use ring:executing-plans (rolling wave: dispatch each> wave — a phase or one epic, your choice — as a workflow → review → user> checkpoint → detail the next phase against the real code → repeat),> or ring:running-dev-cycle for the full subagent-orchestrated workflow.> This document is the living source of truth — task elaboration for later> phases is written back into it during execution.**Goal:** [One sentence describing what this builds]
**Architecture:** [2-3 sentences about approach]
**Tech Stack:** [Key technologies/libraries]
## Phase Overview
| Phase | Milestone | Epics | Status |
|-------|-----------|-------|--------|
| 1 | [what works at the end] | 1.1, 1.2 | Detailed |
| 2 | [what works at the end] | 2.1, 2.2 | Epic-level |
| 3 | [what works at the end] | 3.1 | Epic-level |
---
Epic Format (all phases)
### Epic N.M: [Name]**Goal:** [what exists and works when this epic is done]
**Scope:** [subsystems/directories touched — coarse-grained for later phases]
**Dependencies:** [epics or phases that must land first, or "none"]
**Done when:** [observable acceptance criteria]
**Status:** Pending
**Status:** lifecycle: Pending → Doing → Done | Failed. It is the write target for ring:running-dev-cycle epic checkpoints; ring:executing-plans may ignore it.
For Phase 1 epics, tasks follow immediately below the epic block. For later phases, the epic block is the whole entry — tasks are added during execution.
Task Format (detailed wave only)
Each task is close to a ready-to-dispatch prompt: an implementer with zero context should be able to start within a minute of reading it.
#### Task N.M.T: [Action-oriented name]- [ ] Done
**Context:** [why this task exists; what already exists, with `file.go:42`-style
references into the current codebase]
**Implementation vision:** [the approach; key decisions already made; patterns
to follow or avoid; named edge cases and how each is handled]
**Files:**- Create: `exact/path/to/file.go`- Modify: `exact/path/to/existing.go:123-145`- Test: `path/to/file_test.go`**Verification:** [command to run + expected outcome]
**Done when:** [acceptance criteria]
Use file:line references when pointing into existing code. Paths are always exact for every file touched.
⛔ No Vague Tasks
The plan's deliverable is decisions, not code. A task without decisions is a plan failure:
Pattern
Why it fails
"Add appropriate error handling"
WHICH errors, handled HOW? Decide in the plan.
"Handle edge cases"
Name them, one by one.
"TBD" / "TODO" / "figure out during implementation" in detailed-wave tasks
The detailed wave admits no deferrals — that's what makes it dispatch-ready
Implementation vision that restates the task name
Vision = approach + decisions, not a paraphrase
Task referencing a contract no epic defines
Plan is internally inconsistent
Deferrals ARE allowed in later-phase epics — that is the point of rolling wave. They are NOT allowed inside the detailed wave.
Self-Review
After writing the complete plan, look at the spec with fresh eyes and check the plan against it. This is a checklist you run yourself — not a subagent dispatch.
Check
What to verify
Spec coverage
Skim each requirement in the spec. Point to an epic that covers it. List gaps.
Vagueness scan
Search detailed-wave tasks for the red flags in "No Vague Tasks". Fix any matches.
Contract consistency
Names, signatures, and schemas referenced across epics agree. A contract defined nowhere but used somewhere is a bug.
Phase boundaries
Every phase ends with working, verifiable software.
Verification plausibility
Detailed-wave verification commands target real paths and plausible outcomes.
If you find issues, fix them inline. No need to re-review — just fix and move on.
For high-stakes plans (large surface, multiple authors, critical path): also dispatch a plan-document reviewer subagent using the template in plan-document-reviewer-prompt.md.
Execution Handoff
After saving the plan, offer execution choice:
Plan complete and saved to docs/plans/<filename>.md. Two execution options:
1. Rolling-Wave Execution (this session) — Use ring:executing-plans: you supervise while one workflow per wave-unit (a whole phase, or one epic at a time — you choose at start) implements its tasks; you review what returns, checkpoint, then elaborate the next phase into tasks against the real landed code, and repeat. Best for iterative delivery with course-correction between waves.
2. Subagent-Orchestrated (ring:running-dev-cycle) — lean backend cycle (Gate 0/8/9) with parallel specialist dispatch. Best for production work that must pass through the full review pool.
Which approach?
If Rolling-Wave chosen: Continue with ring:executing-plans in this session.
If Subagent-Orchestrated chosen: Hand off to ring:running-dev-cycle, which owns implementation across the lean cycle (Gate 0 per task, Gates 8/9 per epic, phase boundaries per phase).
Verification Checklist
Before marking the plan complete:
Plan language confirmed with user (prose only; structural tokens, paths, and code stay English)
Plan header present (Goal, Architecture, Tech Stack, Phase Overview)
Every phase ends in working, testable software
Every epic has Goal, Scope, Dependencies, Done-when, Status
Phase 1 epics fully broken into dispatch-ready tasks; later phases epic-level only
No vague tasks in the detailed wave (no "appropriate", "TBD", unnamed edge cases)
Code snippets only where the Code Snippet Policy justifies them
Contract consistency across epics
Self-review checklist applied
Plan saved to docs/plans/YYYY-MM-DD-<feature-name>.md (or docs/pre-dev/{feature}/plan.md when invoked from pre-dev)
Execution handoff offered
Worked Example
### Epic 1.1: Transaction lookup service path
Goal:GET /transactions/:id returns a persisted transaction end-to-end
Scope:internal/service/, internal/handler/Dependencies: none
Done when: integration test fetches a seeded transaction by ID; unknown ID returns 404
Status: Pending
Task 1.1.1: Implement GetTransactionByID service method
Done
Context:TransactionRepository interface already exposes GetByID at internal/domain/repository.go:15. The service layer (internal/service/transaction_service.go) has no read path yet — only Create.
Implementation vision: Add GetByID(ctx, id) to transactionService, delegating to the repository. Follow the observability pattern used by Create (transaction_service.go:31-38): tracking from context, span around the call, HandleSpanError on failure. Propagate domain.ErrNotFound untouched — the handler layer maps it to 404; do not wrap it. No input validation here: ID format is validated at the handler.
Verification:go test ./internal/service/... -run TestTransactionService_GetByID -v — found and not-found cases both pass; not-found asserts errors.Is(err, domain.ErrNotFound).
Done when: service returns the transaction for a known ID and domain.ErrNotFound for an unknown one, with span + log coverage matching the Create pattern.
Epic 2.1: Transaction list endpoint with cursor pagination
Goal:GET /transactions returns tenant-scoped pages of transactions
Scope:internal/service/, internal/handler/, repository query layer
Dependencies: Epic 1.1 (read path patterns established there)
Done when: paginated listing works against seeded data; cursor round-trips; page size capped at 100
Status: Pending
(No tasks yet — elaborated by ring:executing-plans after Phase 1 lands, against the read-path patterns Phase 1 actually established.)
