Plans
Step 0 — Purpose
Ask which format to write — unless the request or briefing already names one.
The implementation plan is the default: a hand-off from think:brainstorming,
planning:executing-plans, agents:subagent-driven-development or the harness gets it
without a question.
| Purpose |
Signal |
Path |
| implementation plan |
code/migration/refactor work, breaking a plan or feature into tickets/issues, any plan hand-off |
Overview → Process below |
| blog |
technical blog, 회고 글, tutorial |
../prd/references/document.md + references/examples/blog.md |
| feedback |
feedback to a colleague, praise, peer review |
references/examples/sbi.md (short form, no section loop) |
Overview
This skill only produces plans — it never runs them. The gap check and
ambiguity check planning:executing-plans would otherwise run at hand-off
get done here instead, at production time: every step gets an observable
check stamped on it before the plan counts as finished, in the same
pass-bar vocabulary that skill's QualityGate judges against. Staleness/
drift stays out of scope on purpose — a fact about when execution
happens, not how the plan was written — so planning:executing-plans owns
that check at hand-off. A clean gap/ambiguity pass here claims nothing
about drift.
Announce at start: "I'm using the plans skill to create the
implementation plan."
Save plans to: docs/plans/YYYY-MM-DD-<feature-name>.md (a stated user
preference overrides this default).
Process
- Scope check. One plan, one subsystem — split multi-subsystem specs first.
- Survey, then structure. Read entry points, tests, and the nearest
analogue before inventing file paths; lock which files are touched and
what each owns.
- Right-size the tasks. One task is one vertical slice: it leaves one
behaviour working end to end at a thinner scope — never "all backend, then
all frontend", because a layer-only task has nothing to demo or test until
its sibling lands. A wide refactor goes expand → migrate → contract, one task
each. Inside a task, 2-5 minute steps: test → fail → implement → pass → commit.
- Gap check, in-line. Confirm each thing a task consumes was produced
by an earlier task — the defect
planning:executing-plans screens for;
catch it before it ships.
- Ambiguity check, in-line. Read each step as a stranger would; if
they'd guess, resolve it now.
- Stamp a pass bar per step. The one observable check proving the step
is done ("that test now passes", "the endpoint returns 429"). No
statable bar means the step isn't finished — go back.
- Self-review. Scan for placeholders ("TBD", "similar to Task N") and
keep names/signatures consistent across tasks.
Output Template
# [Feature Name] Implementation Plan
> Produced by write:plans. Owner for execution routing:
> planning:executing-plans. Steps use checkbox (`- [ ]`) syntax.
**Goal:** [one sentence] **Architecture:** [2-3 sentences]
**Tech Stack:** [key libraries]
---
Per task:
### Task N: [Component]
**Files:** create/modify/test — exact paths.
**Interfaces:** consumes [earlier tasks' signatures] / produces [names
later tasks rely on].
**Blocked by:** [Task k, … | none] — the order made explicit, not implied.
**Pass bar:** [the observable check from Process step 6]
- [ ] 1: failing test (full code) → 2: confirm it fails → 3: minimal
implementation (full code) → 4: confirm it passes → 5: commit
No placeholders where real content belongs: no "TBD", no "similar to Task
N" without the actual code, no reference to a type or function no earlier
task defines.
Dual-Mode
| Mode |
Produces |
Consumed by |
| Solo |
Plan doc above, pass bar per step |
completion:verification-before-completion |
| Harness-engaged |
SetGoal goal-spec — subgoals with acceptance[]/test[] |
The harness QualityGate, subgoal then goal-level |
Compact SetGoal example (≤3 subgoals, ≤6 acceptance criteria total):
{
"goal": "Add rate limiting to the public API",
"acceptance": ["All public endpoints reject over-limit requests with 429"],
"subgoals": [{
"id": "s1",
"title": "Token-bucket limiter middleware",
"skills": ["develop:spring-boot-engineer"],
"acceptance": [
"Requests over the configured rate return 429",
"Requests under the rate pass through unchanged"
],
"test": ["./gradlew test --tests RateLimiterTest"],
"deps": []
}],
"max_retries": 2
}
What Claude Does / What You Do
| Claude |
You |
| Surveys the codebase, locks file structure, right-sizes tasks |
Confirm the subsystem boundary at scope check |
| Runs gap/ambiguity checks in-line, stamps a pass bar per step |
Flag any task that still reads ambiguous |
| Builds the SetGoal goal-spec directly in harness mode |
Route the finished plan onward |
Related Skills
write:prd — PRDs, design docs, RFCs and proposals (moved out of this skill).
write:adr — ADRs and design reviews (moved out of this skill).
write:requirements — split a requirement into items with criteria and open questions.
planning:executing-plans — owns execution routing and the
staleness/drift check this skill leaves out on purpose (downstream).
agents:subagent-driven-development — likely executor once
executing-plans routes a sequential/dependent plan.
completion:verification-before-completion — settles each step's
done-verdict against the pass bar stamped here, in solo mode.
harness:harness — the six-stage engine this skill's harness-mode
output feeds directly as a goal-spec.
1---2name: plans3description: "Use when writing an implementation plan, tickets, tech blog or feedback. Triggers: \"구현 계획 써줘\", \"티켓으로 쪼개줘\", \"기술 블로그\", \"피드백 써줘\", \"write a plan\", \"break into tickets\", \"write a blog\", \"write feedback\""4scenarios:5 - "이 기능 구현 계획 작성해줘"6 - "Create an implementation plan for this new service"7 - "Break this feature into tickets"8 - "Kafka 도입 과정을 기술 블로그로 써줘"9 - "I need to give feedback but don't want it to sound mean"10compatibility:11 optional:12 - sequential-thinking # tracking dependency chains across many tasks13 - think-tool # judging whether a step is unambiguous; which document section has most unknowns14 remote_mcp_note: >-15 sequential-thinking이 있으면 작업 간 의존성 사슬이 복잡한 계획의 갭을 더 체계적으로16 찾을 수 있습니다. Claude 설정 → MCP Servers에서 remote SSE 엔드포인트를 추가하세요.17metadata:18 source: https://skillmd.com/skills/newkayak12/plans19---2021# Plans2223## Step 0 — Purpose2425Ask which format to write — unless the request or briefing already names one.26The implementation plan is the default: a hand-off from `think:brainstorming`,27`planning:executing-plans`, `agents:subagent-driven-development` or the harness gets it28without a question.2930| Purpose | Signal | Path |31|---|---|---|32| implementation plan | code/migration/refactor work, breaking a plan or feature into tickets/issues, any plan hand-off | Overview → Process below |33| blog | technical blog, 회고 글, tutorial | `../prd/references/document.md` + `references/examples/blog.md` |34| feedback | feedback to a colleague, praise, peer review | `references/examples/sbi.md` (short form, no section loop) |3536## Overview3738This skill only produces plans — it never runs them. The gap check and39ambiguity check `planning:executing-plans` would otherwise run at hand-off40get done here instead, at production time: every step gets an observable41check stamped on it before the plan counts as finished, in the same42pass-bar vocabulary that skill's QualityGate judges against. Staleness/43drift stays out of scope on purpose — a fact about *when* execution44happens, not how the plan was written — so `planning:executing-plans` owns45that check at hand-off. A clean gap/ambiguity pass here claims nothing46about drift.4748**Announce at start:** "I'm using the plans skill to create the49implementation plan."5051**Save plans to:** `docs/plans/YYYY-MM-DD-<feature-name>.md` (a stated user52preference overrides this default).5354## Process55561. **Scope check.** One plan, one subsystem — split multi-subsystem specs first.572. **Survey, then structure.** Read entry points, tests, and the nearest58 analogue before inventing file paths; lock which files are touched and59 what each owns.603. **Right-size the tasks.** One task is one vertical slice: it leaves one61 behaviour working end to end at a thinner scope — never "all backend, then62 all frontend", because a layer-only task has nothing to demo or test until63 its sibling lands. A wide refactor goes expand → migrate → contract, one task64 each. Inside a task, 2-5 minute steps: test → fail → implement → pass → commit.654. **Gap check, in-line.** Confirm each thing a task consumes was produced66 by an earlier task — the defect `planning:executing-plans` screens for;67 catch it before it ships.685. **Ambiguity check, in-line.** Read each step as a stranger would; if69 they'd guess, resolve it now.706. **Stamp a pass bar per step.** The one observable check proving the step71 is done ("that test now passes", "the endpoint returns 429"). No72 statable bar means the step isn't finished — go back.737. **Self-review.** Scan for placeholders ("TBD", "similar to Task N") and74 keep names/signatures consistent across tasks.7576## Output Template7778```markdown79# [Feature Name] Implementation Plan8081> Produced by write:plans. Owner for execution routing:82> planning:executing-plans. Steps use checkbox (`- [ ]`) syntax.8384**Goal:** [one sentence] **Architecture:** [2-3 sentences]85**Tech Stack:** [key libraries]86---87```8889**Per task:**9091```markdown92### Task N: [Component]93**Files:** create/modify/test — exact paths.94**Interfaces:** consumes [earlier tasks' signatures] / produces [names95 later tasks rely on].96**Blocked by:** [Task k, … | none] — the order made explicit, not implied.97**Pass bar:** [the observable check from Process step 6]9899- [ ] 1: failing test (full code) → 2: confirm it fails → 3: minimal100 implementation (full code) → 4: confirm it passes → 5: commit101```102103No placeholders where real content belongs: no "TBD", no "similar to Task104N" without the actual code, no reference to a type or function no earlier105task defines.106107## Dual-Mode108109| Mode | Produces | Consumed by |110|---|---|---|111| Solo | Plan doc above, pass bar per step | `completion:verification-before-completion` |112| Harness-engaged | SetGoal goal-spec — subgoals with `acceptance[]`/`test[]` | The harness QualityGate, subgoal then goal-level |113114**Compact SetGoal example** (≤3 subgoals, ≤6 acceptance criteria total):115116```jsonc117{118 "goal": "Add rate limiting to the public API",119 "acceptance": ["All public endpoints reject over-limit requests with 429"],120 "subgoals": [{121 "id": "s1",122 "title": "Token-bucket limiter middleware",123 "skills": ["develop:spring-boot-engineer"],124 "acceptance": [125 "Requests over the configured rate return 429",126 "Requests under the rate pass through unchanged"127 ],128 "test": ["./gradlew test --tests RateLimiterTest"],129 "deps": []130 }],131 "max_retries": 2132}133```134135## What Claude Does / What You Do136137| Claude | You |138|---|---|139| Surveys the codebase, locks file structure, right-sizes tasks | Confirm the subsystem boundary at scope check |140| Runs gap/ambiguity checks in-line, stamps a pass bar per step | Flag any task that still reads ambiguous |141| Builds the SetGoal goal-spec directly in harness mode | Route the finished plan onward |142143## Related Skills144145- `write:prd` — PRDs, design docs, RFCs and proposals (moved out of this skill).146- `write:adr` — ADRs and design reviews (moved out of this skill).147- `write:requirements` — split a requirement into items with criteria and open questions.148- `planning:executing-plans` — owns execution routing and the149 staleness/drift check this skill leaves out on purpose (downstream).150- `agents:subagent-driven-development` — likely executor once151 executing-plans routes a sequential/dependent plan.152- `completion:verification-before-completion` — settles each step's153 done-verdict against the pass bar stamped here, in solo mode.154- `harness:harness` — the six-stage engine this skill's harness-mode155 output feeds directly as a goal-spec.