Writing Plans

Writing Plans

Overview

Write comprehensive implementation plans assuming the engineer has zero context for our codebase and questionable taste. Document everything they need to know: which files to touch for each task, code, testing, docs they might need to check, how to test it. Give them the whole plan as bite-sized tasks. DRY. YAGNI. TDD. Frequent commits.

Assume they are a skilled developer, but know almost nothing about our toolset or problem domain. Assume they don't know good test design very well.

Announce at start: "I'm using the writing-plans skill to create the implementation plan."

Context: This should be run in a dedicated worktree (created by brainstorming skill).

Save plans to: docs/plans/YYYY-MM-DD-<feature-name>.md

Research Before Writing (CRITICAL for per-site fix plans)

Plans that enumerate specific call sites or other concrete code targets must READ the code before constructing the per-site mapping. Reasoning from names and line proximity is reliably wrong.

Before listing per-site fixes, grep the codebase for:

• Existing helpers serving the same purpose. A "new helper" that duplicates an existing one will be caught by the eng critic as DRY violation / reinvention. One grep -r <obvious-name> first — resolveTenantId, requireAuth, withDb, etc. — saves a plan-stage retry.
• The actual call chain into each function being edited. Don't assume from the surface command name. cmdX may call api.x(ctx, id) which calls the primitive; the per-site mapping must target the primitive's actual direct caller, not the surface command the user types.
• Enclosing-function names by line. Don't label a call site from line proximity to a previously-known function ("around line 2700, must be cmdResolve") — grep for function <name> and bound the ranges. Mislabels make execute rediscover the enclosing function instead of following the plan.
• Same-function siblings of in-scope sites that share the same defect class. If your fix touches sites A and B inside function F, audit F for OTHER calls that share A and B's defect class. Deferring same-function siblings while folding their neighbours is a consistency gap the review critics catch.

A plan that skips these is reliably caught at the plan-eng-critic gate and triggers a retry. The grep cost is far below the retry cost. Learned 2026-05-23 from the hippo tenant-isolation residue plan Revision 0, which made all four of these mistakes and failed plan-eng at score 58.

Bite-Sized Task Granularity

Each step is one action (2-5 minutes):

• "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 Claude:** REQUIRED SUB-SKILL: Use executing-plans to implement this plan task-by-task.

**Goal:** [One sentence describing what this builds]

**Architecture:** [2-3 sentences about approach]

**Tech Stack:** [Key technologies/libraries]

---

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`

**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: Write minimal implementation

def function(input):
    return expected

Step 4: Run test to verify it passes

Run: pytest tests/path/test.py::test_name -v Expected: PASS

Step 5: Commit

git add tests/path/test.py src/path/file.py
git commit -m "feat: add specific feature"

## Remember
- Exact file paths always
- Complete code in plan (not "add validation")
- Exact commands with expected output
- Reference relevant skills with @ syntax
- DRY, YAGNI, TDD, frequent commits

## Execution Handoff

After saving the plan, announce: "Plan complete and saved to `docs/plans/<filename>.md`."

Then pick the execution mode yourself — auto-decide, no menu:

- **Default — Subagent-Driven (this session):** fresh subagent per task, review between tasks. REQUIRED SUB-SKILL: subagent-driven-development.
- **Parallel Session** only when the plan is too large for this session's remaining context, or the user asked for a separate session: guide them to open a new session in the worktree. REQUIRED SUB-SKILL: executing-plans.

Name the choice in one line and start. The user can redirect.