Writing Plans
Overview
Write implementation plans for an engineer who has not seen this codebase or this spec. Assume they write idiomatic code in the project's language once they know the exact interface and the exact test, and that they will make a reasonable choice wherever the plan leaves one open. What they cannot know is what you decided: which files, which names and signatures, which values from the spec, which tests prove each task. Document those. Give them the whole plan as bite-sized tasks. DRY. YAGNI. TDD. Frequent commits.
Announce at start: "I'm using the writing-plans skill to create the implementation plan."
Context: If working in an isolated worktree, it should have been created with the host's worktree feature at execution time.
Save plans to: docs/work/YYYY-MM-DD-<topic>/plan.md (next to its spec)
- (User preferences for plan location override this default)
Scope Check
If the spec covers multiple independent subsystems, it should have been broken into sub-project specs during brainstorming. If it wasn't, suggest breaking this into separate plans — one per subsystem. Each plan should produce working, testable software on its own.
File Structure
Before defining tasks, map out which files will be created or modified and what each one is responsible for. This is where decomposition decisions get locked in.
- Design units with clear boundaries and well-defined interfaces. Each file should have one clear responsibility.
- You reason best about code you can hold in context at once, and your edits are more reliable when files are focused. Prefer smaller, focused files over large ones that do too much.
- Files that change together should live together. Split by responsibility, not by technical layer.
- In existing codebases, follow established patterns. If the codebase uses large files, don't unilaterally restructure - but if a file you're modifying has grown unwieldy, including a split in the plan is reasonable.
This structure informs the task decomposition. Each task should produce self-contained changes that make sense independently.
Task Right-Sizing
A task is the smallest unit that carries its own test cycle and is worth a fresh reviewer's gate. When drawing task boundaries: fold setup, configuration, scaffolding, and documentation steps into the task whose deliverable needs them; split only where a reviewer could meaningfully reject one task while approving its neighbor. Each task ends with an independently testable deliverable.
Step Granularity
Each step is one action with a checkable result:
- "Write the failing test" - step
- "Run it to make sure it fails" - step
- "Implement the minimal code to make the test pass" - step
- "Run the tests and make sure they pass" - step
- "Commit" - step
Plan Document Header
Every plan MUST start with this header:
# [Feature Name] Implementation Plan
> **For agentic workers:** Implement this plan task-by-task in the current session (Native), then dispatch a fresh `evaluator` agent to review the whole branch. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** [One sentence describing what this builds]
**Architecture:** [2-3 sentences about approach]
**Tech Stack:** [Key technologies/libraries]
**Spec:** [path to the spec/design doc this plan implements — the plan
argues from the spec, so the spec travels with it; executors read both]
## Global Constraints
[The spec's project-wide requirements — version floors, dependency limits,
naming and copy rules, platform requirements — one line each, with exact
values copied verbatim from the spec. Every task's requirements implicitly
include this section.]
## Review Focus
[The five input classes or failure modes the spec implies but no task's
tests exercise that are most likely to bite a person using this software
— one line each, naming the input or condition and the behavior a
reasonable person would expect, most likely first. The spec is a vision
document: it says what the software must do, not everything it will
meet, and its silence on an input is not permission for that input to
break the program. Write the list here, once, with the spec in front of
you. Then, for each line, add the test that pins it to the task that
owns the code, in that task's own step style.]
---
Task Structure
### Task N: [Component Name]
**Files:**
- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py:123-145`
- Test: `tests/exact/path/to/test.py`
**Interfaces:**
- Consumes: [what this task uses from earlier tasks — exact signatures]
- Produces: [what later tasks rely on — exact function names, parameter
and return types. A task's implementer sees only their own task; this
block is how they learn the names and types neighboring tasks use.]
- [ ] **Step 1: Write the failing test**
```python
def test_specific_behavior():
result = function(input)
assert result == expected
```
- [ ] **Step 2: Run test to verify it fails**
Run: `pytest tests/path/test.py::test_name -v`
Expected: FAIL with "function not defined"
- [ ] **Step 3: Implement `function(input: InputType) -> ResultType` in `exact/path/to/file.py`**
One line on the approach when the signature and the test leave a choice
(which library call, which data structure); a code block only for an
algorithm they do not determine.
- [ ] **Step 4: Run test to verify it passes**
Run: `pytest tests/path/test.py::test_name -v`
Expected: PASS
- [ ] **Step 5: Commit**
```bash
git add tests/path/test.py src/path/file.py
git commit -m "feat: add specific feature"
```
What a Step Contains
A step is done when the implementer can write exactly one reasonable thing from it. That is the whole requirement: unambiguous, not complete. Each kind of step carries what makes it unambiguous and nothing more:
- A test step: the test's name and its assertions, as code, with the spec's exact values in them.
- A code step: the exact signature (name, parameters, return type), the file it lives in, and the specific values the spec pins. The implementer writes the body. A body appears only for an algorithm the signature and tests do not determine, or for exact copy the spec fixes.
- A verification step: the command to run and the output that means it passed.
- A reference to another task: that task's Interfaces block says what to use; the plan does not repeat that task's code.
A plan is the set of decisions the implementer cannot make alone. A plan longer than the code it describes has written the code instead. Lines that decide nothing ("TBD", "handle edge cases", "add appropriate validation", "write tests for the above", a type or function no task defines) are the opposite failure, and the self-review catches both.
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.
1. Spec coverage: Skim each section/requirement in the spec. Can you point to a task that implements it? List any gaps.
2. Step scan: Every step must let the implementer write exactly one reasonable thing, and no step may carry more than that: a line that decides nothing is a gap, a function body the signature and tests already determine is a transcript. Fix both.
3. Type consistency: Do the types, method signatures, and property names you used in later tasks match what you defined in earlier tasks? A function called clearLayers() in Task 3 but clearFullLayers() in Task 7 is a bug.
4. Review Focus: For each input class or failure mode the spec implies, is there a task whose tests exercise it? The five uncovered ones most likely to bite a person go in the Review Focus section, and each line there gets its test added to the owning task. An empty section means you checked and found none, not that you skipped the check.
5. Proportion: Compare the plan's length to the spec's. A plan several times longer than the spec it implements is a transcript of the program, not a plan. If code blocks are most of the document, replace bodies with signatures, test names and assertions, and check that each step is still unambiguous.
If you find issues, fix them inline. No need to re-review — just fix and move on. If you find a spec requirement with no task, add the task.
Execution Handoff
After saving and self-reviewing the plan, link it for your human partner to read. If they have already explicitly supplied an execution method, ask them to review the plan and confirm it captures what they want; wait for that review before implementation, then use the preserved method. Otherwise, ask them to review the plan and choose an execution method before implementation.
When no execution method has already been supplied:
"Plan complete and saved to docs/work/<folder>/plan.md. Please review the plan. Which execution approach would you prefer?
- Subagent-driven - A fresh subagent implements each task and a fresh reviewer checks it before the next one starts, then a whole-branch review at the end. Most thorough; costs a fresh context per task and per review.
- Native - I implement every task myself in this session, the way this harness runs work, then one fresh reviewer on the most capable model checks the whole branch. Cheapest and fastest; no independent review until the end. Runs well with a mid-tier session model, since the plan carries the design.
For this plan I recommend , because <one sentence from the plan: how much the tasks depend on each other's interfaces, how many there are, what a shipped mistake would cost>. Does the plan capture what you want, and which approach should we use?"
When an execution method has already been supplied:
"Plan complete and saved to docs/work/<folder>/plan.md. Please review the plan. Does it capture what you want?"
If Subagent-driven chosen:
- Dispatch one fresh subagent per task with that task's section of the plan; after each, a fresh reviewer checks it against the plan before the next task starts. Finish with the whole-branch review below.
If Native chosen:
- Implement the tasks in order in this session, checking off each step and running its verification command. Then dispatch a fresh
evaluatoragent to review the whole branch against the spec and this plan.
이 하네스에서
- 저장 위치: spec과 같은 폴더의
docs/work/YYYY-MM-DD-<주제>/plan.md. - 각 Task 머리에 그 Task가 채우는 spec의 수용 기준 번호(R1…)를 적는다. 어느 Task에도 없는 R#이 있으면 plan을 끝내지 않는다. 화면 Task는 따를
design.md의 절을 적고, 검증에 스크린샷 비교를 넣는다. - 헤더의 Goal 위에 두 줄을 둔다:
승인:(사용자가 승인하면 날짜를 적고 커밋한다)과재개: 첫 미체크 단계부터. 체크된 단계가 만든 파일·테스트가 실제로 있는지 먼저 확인한다.승인:날짜가 없는 plan은 실행하지 않는다. - 구현 중에는 handoff.md
지금:에 plan 경로와 진행 중인 Task를 적고, 단계를 끝낼 때마다 체크박스를 체크해 같은 커밋에 넣는다. 압축·재시작·다른 호스트에서도 그 체크에서 이어 간다. - 구현이 계획과 달라지면 같은 커밋에서 plan.md를 고친다. 요구 자체가 바뀌었으면 brainstorming의 "이 하네스에서"대로 앞 문서부터 고친다.
