Planning and task breakdown
When to use
- Spec or requirements exist but work is too large to start
- Need dependency order, checkpoints, or parallelization notes
When not to use
- Obvious single-file change with clear acceptance criteria
Process
- Read-only plan mode — map codebase patterns; no code yet
- Dependency graph — foundations before dependents
- Vertical slices — prefer end-to-end thin slices over horizontal layers
- Write tasks — each with acceptance + verify + files + size
- Checkpoints — every 2-3 tasks, full verify
High-quality plan checklist
- Each task is small enough to complete in one focused session (roughly 2–15 minutes of agent work)
- Files: lists exact paths to create or edit (no "update relevant files")
- Verify: includes full command and expected outcome (exit code, key output line)
- No placeholders like "add tests here" or "implement logic"
- For behavior changes, note which test should fail first (TDD) when applicable
Task template
## Task N: [Title]
**Acceptance:** ...
**Verify:** `command` -> expected result
**Depends on:** ...
**Files:** path/to/file.ext, path/to/test.ext
**Size:** S | M | L (L = split further)
Sizing
| Size | Files | Guidance |
|---|---|---|
| S | 1-2 | One endpoint or component |
| M | 3-5 | One feature slice |
| L | 5+ | Split before implementing |
Parallelization
- Safe parallel: independent slices, docs, tests for stable code
- Sequential: migrations, shared contracts, shared state
Output
Plan doc with phases, risks, open questions. Human approval before implementation.
Related
spec-driven-development, incremental-implementation, verify-before-done
Plan quality patterns inspired by obra/superpowers (MIT).