| name | tgd-planning-and-task-breakdown |
| description | Breaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. |
Planning and Task Breakdown
Overview
Decompose work into small, verifiable tasks with explicit acceptance criteria. Good task breakdown is the difference between an agent that completes work reliably and one that produces a tangled mess. Every task should be small enough to implement, test, and verify in a single focused session.
When to Use
- You have a spec and need to break it into implementable units
- A task feels too large or vague to start
- Work needs to be parallelized across multiple agents or sessions
- You need to communicate scope to a human
- The implementation order isn't obvious
When NOT to use: Single-file changes with obvious scope, or when the spec already contains well-defined tasks.
The Zero-Context Rule
Plans must assume the executor has:
- Zero context about the codebase
- Limited knowledge of the problem domain
- Questionable taste in design decisions
- Weak test design skills
This means every task must contain:
- Exact file paths (not "the config file")
- Exact commands to run (not "run the tests")
- Code blocks with actual implementation hints (not "implement the function")
- Expected outputs (not "verify it works")
Never write:
- "Similar to Task N" (tasks may be read out of order)
- "TBD" or "TODO" or "fill in details"
- "Add appropriate error handling" (specify which errors, which handling)
- Steps that describe what without showing how
The Planning Process
Step 0: Feature Name Resolution
Before planning, determine and validate <feature-name>:
- Verify: Scan
$TGD_DIR/ for feature directories — subdirectories containing SPEC.md or PRD.md. Infrastructure dirs (.scans/, wiki/, and any dot-directories) are NOT features — always exclude them. If exactly one feature directory exists, lock that name.
- Ask (if ambiguous): If multiple feature directories exist, list them and ask the user to pick. If none exist, STOP — run
/tgd-define first.
- Lock: All planning artifacts go into
$TGD_DIR/<feature-name>/.
Step 1: Enter Plan Mode (Read-Only Analysis)
Before writing any code, operate in read-only mode to gather context from all available tGD artifacts:
- Read
$TGD_DIR/CONTEXT.md: Understand existing project structure, tech stack, and conventions.
- Read
$TGD_DIR/<feature-name>/PRD.md: Understand the business goals, user pain points, and scope boundaries.
- Read
$TGD_DIR/<feature-name>/SPEC.md: Analyze technical requirements, API contracts, and database schemas.
- Read
$TGD_DIR/<feature-name>/DESIGN.md (if present): Review component trees and UI flows.
Synthesis: Map dependencies between existing code and new requirements. Note risks and unknowns. If .codegraph/ exists, run codegraph impact "<core-symbol>" on any symbol the feature will modify to assess blast radius and inform task ordering. If planning a large refactor, run the understand-diff skill to visualize the impact of proposed changes before breaking down tasks.
Do NOT write code during planning. Write a plan document at $TGD_DIR/<feature-name>/TASKS.md covering: dependency graph, ordered task list with acceptance criteria, verification checkpoints, and risks with mitigations.
TASKS.md template (save to $TGD_DIR/<feature-name>/TASKS.md):
# TASKS.md: [Feature Name]
> **Corresponding PRD**: [PRD.md](PRD.md)
> **Tech Stack**: [List from SPEC]
## Overview
[One paragraph summary of what we're building]
## Architecture Decisions
- [Key decision 1 and rationale]
- [Key decision 2 and rationale]
---
## Task 1: [User Story Title] (Story ID: US-01)
**Status:** pending <!-- pending | in-progress | complete | blocked: <ref>. The ONLY task-state marker: /tgd-develop flips it, resume + re-plan read it. Do not invent checkboxes/emoji variants. -->
**Spec-Review:** pending <!-- pending | PASS — <one line> | FAIL — <one line>. Flipped by /tgd-develop's two-stage review; /tgd-verify fails closed on a complete task left pending. -->
**Quality-Review:** pending <!-- same states; flipped after the code-quality review pass. -->
### 1. Context & Goal
[What is the goal of this task? Why is it important?]
- **Priority**: [High/Medium/Low]
- **Dependencies**: [None / Task N]
### 2. Technical Design
**Database Schema (if any):**
```[Language]
// Example: Prisma Schema or SQL
API Contract:
- Method
/endpoint
- Input:
{ ... }
- Output:
Status Code { ... }
3. Acceptance Criteria (BDD)
-
Every task must use BDD (Given/When/Then) format — this ensures all criteria are behavior-level, testable, and consistent with REGRESSION-CATALOG entries.
-
Every criterion carries a stable ID: AC-<task>.<n> (AC-1.1, AC-1.2, …). The verifying test MUST mention this ID in its name, docstring, or a comment — ac-trace.py cross-references them during /tgd-verify.
-
AC-1.1 — Given [initial context] When [event happens] Then [expected outcome]
- Regression: [Yes
[R] / No]
- Test: [
tests/path/to/test.ts — filled during /tgd-develop; MANDATORY for [R] criteria]
-
Every criterion declares its carrier — Test: for anything with runtime
behavior, or Doc: for documentation-only criteria (README sections, usage
examples — content with no runtime surface):
- Doc:
README.md contains "getMonthlySummary("
ac-trace.py verifies a Doc: carrier by checking the named file exists and
contains the quoted string — no test reference is required or expected.
Without this, doc-only criteria force a workaround (agents paste the AC id
into an unrelated test file as a comment just to satisfy the trace).
Doc-only criteria can never be [R] — the regression catalog replays
executable tests only.
-
[R] marking rules — must mark [R] if the criterion matches ANY of:
- (a) Verifies an acceptance criterion from the PRD's User Stories table (a US-xx row's "Acceptance Criteria" column) or a PRD Success Metric. (Note the namespaces: PRD rows are
US-xx; the AC-<task>.<n> ids exist only in TASKS.md.)
- (b) Covers a critical user path (auth, payment, data loss, security boundary)
- (c) Catches a previously-fixed bug from
REGRESSION-CATALOG.md
- SHOULD NOT mark
[R] if criterion is: cosmetic, internal refactor, dev-only tooling, single-use migration
- When in doubt: mark
[R]. The cost of an extra catalog entry is low; the cost of missing a regression is high.
- If
[R]: a corresponding test MUST be created during /tgd-develop (TDD) and its path recorded in the criterion's Test: field. It will be added to $TGD_DIR/REGRESSION-CATALOG.md during /tgd-release.
- Enforcement (machine-gated):
/tgd-verify runs python3 $TGD_REPO_ROOT/scripts/ac-trace.py $TGD_DIR/<feature>/ <client-repo> — it fails when any AC id is unreferenced by tests (or, for Doc: criteria, when the named file is missing or lacks the quoted string), when any [R] criterion lacks a Test: file, or when that file is missing on disk. A TASKS.md without AC ids fails closed.
4. Files Likely Touched
path/to/file.tstests/path/to/test.ts
Checkpoint: Verification
✅ All tests pass (npm test)
✅ Build succeeds
✅ Lint clean
Risks & Mitigations
| Risk | Impact | Mitigation |
|---|
| [Risk] | High/Med/Low | [Strategy] |
Open Questions
- [Question needing human input]
Sign-off
**This is the ONLY TASKS.md format.** Everything below describes how to fill
it in — do not invent alternative task layouts; `ac-trace.py` (run by
`/tgd-verify`) fails closed on task lists without `AC-<task>.<n>` ids.
### Step 2: Identify the Dependency Graph
Map what depends on what:
Database schema
│
├── API models/types
│ │
│ ├── API endpoints
│ │ │
│ │ └── Frontend API client
│ │ │
│ │ └── UI components
│ │
│ └── Validation logic
│
└── Seed data / migrations
Implementation order follows the dependency graph bottom-up: build foundations first.
### Step 3: Slice Vertically
Instead of building all the database, then all the API, then all the UI — build one complete feature path at a time:
**Bad (horizontal slicing):**
Task 1: Build entire database schema
Task 2: Build all API endpoints
Task 3: Build all UI components
Task 4: Connect everything
**Good (vertical slicing):**
Task 1: User can create an account (schema + API + UI for registration)
Task 2: User can log in (auth schema + API + UI for login)
Task 3: User can create a task (task schema + API + UI for creation)
Task 4: User can view task list (query + API + UI for list view)
Each vertical slice delivers working, testable functionality.
### Step 4: Write Tasks
Write each task using the **canonical per-task block from the TASKS.md
template above** (`**Status:** pending` line → Context & Goal → Technical
Design → Acceptance Criteria with `AC-<task>.<n>` ids, `[R]` decision, and
`Test:` field → Files Likely Touched). Do not use a simplified layout — the
AC ids and `Test:` fields are machine-checked downstream, and the `Status:`
line is what `/tgd-develop`'s resume rule and `/tgd-plan`'s re-plan protocol
read; a task without one is invisible to both.
Per-task quality bar (in addition to the template fields):
- **Verification is explicit**: name the exact command (`npm test -- --grep "x"`),
not "run the tests" (Zero-Context Rule)
- **Dependencies declared**: task numbers this depends on, or "None"
- **Scope estimated**: Small (1-2 files) / Medium (3-5) / Large (5+ → split it)
- **No standalone "write tests" tasks**: tests are each task's AC carriers,
written inside that task's TDD cycle — never a separate task scheduled after
the implementation. A "Task N: write the test suite" decomposition
structurally encourages test-after (the implementation task completes
untested, then the test task rubber-stamps it) and produces meta-criteria
("tests exist and pass") that trace to the same test lines as the behavior
criteria they duplicate. If a task's tests feel like a separate work item,
the task is too big — split the task, not the testing.
### Step 5: Order and Checkpoint
Arrange tasks so that:
1. Dependencies are satisfied (build foundation first)
2. Each task leaves the system in a working state
3. Verification checkpoints occur after every 2-3 tasks
4. High-risk tasks are early (fail fast)
Add explicit checkpoints:
```markdown
## Checkpoint: After Tasks 1-3
- [ ] All tests pass (`npm test`)
- [ ] Application builds without errors (`npm run build`)
- [ ] Core user flow works end-to-end
Checkpoint items must be machine-checkable commands — /tgd-develop runs
them when it reaches the checkpoint and stops only on failure; it does NOT
pause for a human. Human review is concentrated in /tgd-release's sign-off
gate, not mid-execution.
Instrumentation Tasks and TRACKING-PLAN.md
If the PRD's §6 Success Metrics names a tracking event that does not exist yet (rule 2 in tgd-spec-driven-development's §6 filling rules), planning owns two outputs:
1. Register the event in $TGD_DIR/TRACKING-PLAN.md — the cumulative event dictionary shared across ALL features and ALL repos of the project (same role as REGRESSION-CATALOG.md: append-only, one source of truth). Create the file on first use with this header, then append one entry per event:
# Tracking Plan
> Cumulative event dictionary across all features and platforms.
> Event names are semantic and platform-agnostic — see the three rules below.
---
### sign_up_completed
- **Semantic trigger:** User completes registration (server returns 201)
- **Source of truth:** server
- **Platforms:** server (web/ios do NOT duplicate this event)
- **Properties:** `method: "email"|"oauth"` · `platform` (auto-attached by SDK) · no PII
- **Feature:** user-login
- **Status:** planned <!-- flipped to "live since vYYYY.MM.DD" by /tgd-release -->
Cross-platform rules (non-negotiable):
- One semantic = one event name; platform is a property, not a suffix.
sign_up_completed + platform: web|ios|android — never sign_up_web / signUpIos triplets. Split names silently break every funnel that forgets one variant.
- Triggers are defined semantically, not by UI. "Registration completed = server returns 201", not "register button clicked" — each platform's UI differs; the semantic moment is the single shared definition every implementer maps to.
- Every event declares a source of truth. Conversion-critical events (signup, payment) default to server-side — one implementation covers all platforms and survives ad-blockers. Pure interaction events (scroll, hover) are client-side. Often the answer to "do all three platforms need this?" is "no — server emits it once."
Naming convention: object_action in snake_case, property keys also snake_case — on every platform. Payload key drift (plan_type vs planType) is exactly what the dictionary exists to prevent.
2. Create one instrumentation task per platform listed in the entry's Platforms field — a normal task in TASKS.md (multi-repo tagged if applicable) with its own BDD acceptance criteria, e.g.:
- **AC-4.1** — **Given** a user completes registration **When** the server responds 201 **Then** `sign_up_completed` is emitted with properties `method`, `platform` and no PII
- **Regression**: No
- **Test**: [filled during /tgd-develop — the test asserts the event fires with the expected payload keys]
A mis-firing event is worse than a missing one — you will trust a wrong number. That is why instrumentation gets tested ACs, not a "remember to add analytics" checklist line.
Re-planning an Existing TASKS.md
When TASKS.md already exists (spec changed mid-flight, scope grew), the default is an incremental update, never a from-scratch rewrite — /tgd-develop backfills Test: fields and completion state into TASKS.md, and regenerating the file destroys them, breaking ac-trace.py and the regression chain downstream.
Incremental rules:
- Completed tasks are immutable: any task with
**Status:** complete keeps its Status line, criteria, and Test: fields byte-for-byte. (The Status: lines are also how the re-plan prompt counts "M 個已完成" — no marker, no count.)
- Never renumber existing
AC-<task>.<n> ids — tests already reference them.
- New tasks continue the numbering after the highest existing task.
- An unstarted task invalidated by the spec change may be rewritten in place (same task number, new content, new criterion ids under it).
- A full rewrite is legitimate only when the old plan is void (feature repurposed) AND the user explicitly chose it (
/tgd-plan's re-plan prompt, option 2).
Task Sizing Guidelines
| Size | Files | Scope | Example |
|---|
| XS | 1 | Single function or config change | Add a validation rule |
| S | 1-2 | One component or endpoint | Add a new API endpoint |
| M | 3-5 | One feature slice | User registration flow |
| L | 5-8 | Multi-component feature | Search with filtering and pagination |
| XL | 8+ | Too large — break it down further | — |
If a task is L or larger, it should be broken into smaller tasks. An agent performs best on S and M tasks.
When to break a task down further:
- It would take more than one focused session (roughly 2+ hours of agent work)
- You cannot describe the acceptance criteria in 3 or fewer bullet points
- It touches two or more independent subsystems (e.g., auth and billing)
- You find yourself writing "and" in the task title (a sign it is two tasks)
Parallelization Opportunities
When multiple agents or sessions are available:
- Safe to parallelize: Independent feature slices, tests for already-implemented features, documentation
- Must be sequential: Database migrations, shared state changes, dependency chains
- Needs coordination: Features that share an API contract (define the contract first, then parallelize)
Common Rationalizations
| Rationalization | Reality |
|---|
| "I'll figure it out as I go" | That's how you end up with a tangled mess and rework. 10 minutes of planning saves hours. |
| "The tasks are obvious" | Write them down anyway. Explicit tasks surface hidden dependencies and forgotten edge cases. |
| "Planning is overhead" | Planning is the task. Implementation without a plan is just typing. |
| "I can hold it all in my head" | Context windows are finite. Written plans survive session boundaries and compaction. |
Red Flags
- Starting implementation without a written task list
- Tasks that say "implement the feature" without acceptance criteria
- No verification steps in the plan
- All tasks are XL-sized
- No checkpoints between tasks
- Dependency order isn't considered
Verification
Before starting implementation, confirm:
