andretauan/tao-plan-writing

TAO Plan Writing

When to use

Auto-loaded when creating or updating PLAN.md, decomposing features into tasks, or planning phases.

Task Decomposition Principles

1. One task = one commit — if it needs multiple commits, split it
2. Acceptance criteria are testable — no vague "works correctly"
3. Tasks are ordered by dependency — blocked tasks come after their blockers
4. Each task has estimated effort — S (< 30 min), M (30-90 min), L (90+ min)

PLAN.md Structure

# Phase XX — [Phase Title]

**Goal:** One sentence describing what this phase achieves
**Prerequisite:** Phase XX-1 completed (or "none")

## Tasks

### T01 — [Task Name]
- **Scope:** What files/modules this touches
- **Acceptance criteria:**
  - [ ] Criterion 1 (testable)
  - [ ] Criterion 2 (testable)
- **Effort:** S | M | L
- **Model:** Sonnet | Opus | GPT-4.1
- **Dependencies:** none | T0X

### T02 — [Task Name]
...

Model Routing Rules

| Task Type | Model | Cost | |-----------|-------|------| | Simple CRUD, config, boilerplate | GPT-4.1 | Free | | Standard logic, tests, integration | Sonnet | Low | | Complex architecture, security, debugging | Opus | High |

Decomposition Heuristics

• Database first — schema before code that reads it
• Interface first — API contract before implementation
• Tests alongside — test task follows immediately after implementation task
• Config last — environment/deployment config after all code

Red Flags (split the task if...)

• Task touches more than 3 files
• Task description uses "and" more than once
• Task scope crosses module boundaries
• Estimated effort is L — usually splittable

Quality Gate

Before marking PLAN.md complete, verify:

• [ ] All tasks have acceptance criteria
• [ ] All tasks have effort estimates
• [ ] All tasks have model assignment
• [ ] Dependencies form a DAG (no circular)
• [ ] Task count is between 3-10 per phase