bee:pre-dev-task-breakdown

Task Breakdown - Value-Driven Decomposition

Foundational Principle

Every task must deliver working software that provides measurable user value.

Creating technical-only or oversized tasks creates:

• Work that doesn't ship until "everything is done"
• Teams working on pieces that don't integrate
• No early validation of value or technical approach
• Waterfall development disguised as iterative process

Tasks answer: What working increment will be delivered? Tasks never answer: How to implement that increment (that's Subtasks).

Mandatory Workflow

Phase	Activities
1. Task Identification	Load PRD (Gate 1, required), TRD (Gate 3, required); optional: Feature Map, API Design, Data Model, Dependency Map; identify value streams
2. Decomposition	Per component/feature: define deliverable, set success criteria, map dependencies, estimate effort via AI analysis (max 16 AI-agent-hours), plan testing, identify risks
3. Gate 7 Validation	All TRD components covered; every task delivers working software; measurable success criteria; correct dependencies; no task >2 weeks; testing strategy defined; risks with mitigations; delivery sequence optimizes value

Explicit Rules

✅ DO Include in Tasks

Task ID, title, type (Foundation/Feature/Integration/Polish), deliverable (what ships), user value (what users can do), technical value (what it enables), success criteria (testable/measurable), dependencies (blocks/requires/optional), effort estimate (AI-agent-hours with confidence), testing strategy, risk identification with mitigations, Definition of Done checklist

❌ NEVER Include in Tasks

Implementation details (file paths, code examples), step-by-step instructions (those go in subtasks), technical-only tasks with no user value, tasks exceeding 2 weeks (break them down), vague success criteria ("improve performance"), missing dependency information, undefined testing approach

Task Sizing Rules

Size	AI-agent-hours	Calendar Duration*	Scope
Small (S)	1-4h	1-2 days	Single component
Medium (M)	4-8h	2-4 days	Few dependencies
Large (L)	8-16h	1-2 weeks	Multiple components
XL (>16h)	BREAK IT DOWN	Too large	Not atomic

*Calendar duration assumes 1.5x multiplier (standard validation), 90% capacity, and 1 developer

Value Delivery Rules

• Foundation: Enables other work (database setup, core services)
• Feature: Delivers user-facing capabilities
• Integration: Connects to external systems
• Polish: Optimizes or enhances (nice-to-have)

AI-Assisted Time Estimation

See shared-patterns/ai-agent-baseline.md for baseline definition.

After defining task scope and success criteria, the system automatically estimates implementation time.

Estimation Process

1. Tech Stack Detection: Identify project type from TRD

   • PHP/Laravel → dispatch bee:backend-engineer-php
   • TypeScript Backend → dispatch bee:backend-engineer-typescript
   • React/Next.js → dispatch bee:frontend-engineer
   • Mixed/Unknown → dispatch bee:codebase-explorer

2. Scope Analysis: Specialized agent analyzes:

   • Technical components (endpoints, schemas, services)
   • Complexity level (CRUD vs custom algorithms)
   • Available libraries (lib-commons, frameworks)
   • Test requirements (coverage %, test types)
   • Documentation requirements

3. Time Calculation: Agent estimates per scope item

   • Database schema + migrations
   • Repository/data layer
   • Business logic layer
   • API handlers/controllers
   • Validation logic
   • Error handling
   • Unit tests (TDD)
   • Integration tests
   • Documentation

4. Output: Total in AI-agent-hours

For detailed baseline definition and capacity explanation, see shared-patterns/ai-agent-baseline.md.

Confidence Levels

Level	Criteria	Example
High	Standard patterns, libs available, clear scope	CRUD API with lib-commons
Medium	Some custom logic, partial lib support	Payment integration
Low	Novel algorithms, no lib support, vague scope	ML feature, R&D work

Example Output

**Effort Estimate:**
- **Baseline:** AI Agent via bee:dev-cycle
- **AI Estimate:** 4.5 AI-agent-hours
- **Estimation Method:** bee:backend-engineer-php analysis
- **Confidence:** High (standard CRUD, lib-commons available)

**Breakdown:**
- Database schema + migrations: 0.5h
- Repository layer (CRUD): 0.5h
- Service layer (business logic): 0.5h
- HTTP handlers (4 endpoints): 1.0h
- Input validation: 0.3h
- Error handling: 0.2h
- Unit tests (TDD, 85% coverage): 0.8h
- Integration tests: 0.5h
- OpenAPI documentation: 0.2h
**Total: 4.5 AI-agent-hours**

**Assumptions:**
- lib-commons/http, lib-commons/postgres, lib-commons/validator available
- Standard CRUD patterns (no complex algorithms)
- PostgreSQL database configured
- bee:dev-cycle will execute implementation

**Team Type:** Backend Engineer (PHP)

Rationalization Table

Excuse	Reality
"This 3-week task is fine"	Tasks >2 weeks hide complexity. Break it down.
"Setup tasks don't need value"	Setup enables value. Define what it enables.
"Success criteria are obvious"	Obvious to you ≠ testable. Document explicitly.
"Dependencies will be clear later"	Later is too late. Map them now.
"We don't need detailed estimates"	Without estimates, no planning possible. Size them.
"Technical tasks can skip user value"	Even infrastructure enables users. Define the connection.
"Testing strategy can be decided during"	Testing affects design. Plan it upfront.
"Risks aren't relevant at task level"	Risks compound across tasks. Identify them early.
"DoD is the same for all tasks"	Different tasks need different criteria. Specify.
"We can combine multiple features"	Combining hides value delivery. Keep tasks focused.
"Skip AI estimation, use story points"	Story points are abstract, AI hours are concrete
"Manual estimate is faster"	Fast ≠ accurate. AI analyzes full scope consistently
"AI estimate too low, inflate it"	Inflation happens in multiplier (Gate 9), not here
"Confidence is always High"	Confidence reflects scope clarity and complexity
"Skip breakdown, just give total"	Breakdown enables validation and learning
"AI can't estimate this, too complex"	Complex = lower confidence, not impossible

Red Flags - STOP

If you catch yourself writing any of these in a task, STOP:

• Task estimates over 2 weeks
• Tasks named "Setup X" without defining what X enables
• Success criteria like "works" or "complete" (not measurable)
• No dependencies listed (every task depends on something)
• No testing strategy (how will you verify?)
• "Technical debt" as a task type (debt reduction must deliver value)
• Vague deliverables ("improve", "optimize", "refactor")
• Missing Definition of Done

When you catch yourself: Refine the task until it's concrete, valuable, and testable.

Gate 7 Validation Checklist

Category	Requirements
Task Completeness	All TRD components have tasks; all PRD features have tasks; each task appropriately sized (no XL+); task boundaries clear
Delivery Value	Every task delivers working software; user value explicit; technical value clear; sequence optimizes value
Technical Clarity	Success criteria measurable/testable; dependencies correctly mapped; testing approach defined; DoD comprehensive
Team Readiness	Skills match capabilities; estimates realistic; capacity available; handoffs minimized
Risk Management	Risks identified per task; mitigations defined; high-risk tasks scheduled early; fallback plans exist
Multi-Module (if applicable)	All tasks have target: field; all tasks have working_directory:; per-module files generated (if doc_organization: per-module)

Gate Result: ✅ PASS → Subtasks | ⚠️ CONDITIONAL (refine oversized/vague) | ❌ FAIL (re-decompose)

Multi-Module Task Tagging

If TopologyConfig exists in research.md frontmatter (from Gate 0):

Read Topology Configuration

# From research.md frontmatter
topology:
  scope: fullstack
  structure: monorepo | multi-repo
  modules:
    backend:
      path: packages/api
      language: php
    frontend:
      path: packages/web
      framework: nextjs
  doc_organization: unified | per-module

Task Target Assignment

Each task MUST have target: and working_directory: fields when topology is multi-module.

Agent assignment depends on both target and api_pattern:

Target	API Pattern	Task Type	Agent
backend	any	API endpoints, services, data layer, CLI	bee:backend-engineer-php
backend	any	API endpoints, services, data layer, CLI	bee:backend-engineer-typescript
frontend	direct	UI components, pages, forms, Server Components	bee:frontend-engineer
frontend	direct	Server Actions, data fetching hooks	bee:frontend-engineer
frontend	bff	API routes, data aggregation, transformation	bee:frontend-bff-engineer-typescript
frontend	bff	UI components, pages, forms	bee:frontend-engineer
shared	any	CI/CD, configs, docs, cross-module utilities	DevOps or general

How to Determine Agent for Frontend Tasks

Read api_pattern from research.md frontmatter:

# From research.md
topology:
  scope: fullstack
  api_pattern: direct | bff | other

Decision Flow:

Is task target: frontend?
├─ NO → Use backend-engineer-* based on language
└─ YES → Check api_pattern
    ├─ direct → ALL frontend tasks use frontend-engineer
    └─ bff → Split tasks:
        ├─ API routes, aggregation, transformation → frontend-bff-engineer-typescript
        └─ UI components, pages, forms → frontend-engineer

Task Format with Agent Assignment

## T-003: User Login API Endpoint

**Target:** backend
**Working Directory:** packages/api
**Agent:** bee:backend-engineer-php

**Deliverable:** Working login API that validates credentials and returns JWT token.

...rest of task...

## T-004: User Dashboard Data Aggregation

**Target:** frontend
**Working Directory:** packages/web
**Agent:** bee:frontend-bff-engineer-typescript  # Because api_pattern: bff

**Deliverable:** BFF endpoint that aggregates user profile, recent activity, and notifications.

...rest of task...

## T-005: User Dashboard UI

**Target:** frontend
**Working Directory:** packages/web
**Agent:** bee:frontend-engineer  # UI task, even with BFF pattern

**Deliverable:** Dashboard page component consuming aggregated data from BFF.

...rest of task...

Validation for Agent Assignment

Check	Requirement
All tasks have Agent: field	MANDATORY
Agent matches api_pattern rules	If frontend + bff → check task type
BFF tasks clearly separated	Data aggregation vs UI clearly split
No mixed responsibilities	One task = one agent

Per-Module Output

Document placement depends on topology.structure:

Single-Repo

All tasks in one file:

docs/pre-dev/{feature}/
└── tasks.md           # All tasks with target tags

Monorepo (per-module placement)

Index at root, filtered tasks in module directories:

docs/pre-dev/{feature}/
└── tasks.md           # Index with ALL tasks (target tags included)

{backend.path}/docs/pre-dev/{feature}/
└── tasks.md           # Backend tasks only (target: backend)

{frontend.path}/docs/pre-dev/{feature}/
└── tasks.md           # Frontend tasks only (target: frontend)

Multi-Repo (distributed placement)

Tasks distributed to respective repositories:

{backend.path}/docs/pre-dev/{feature}/
└── tasks.md           # Backend tasks only

{frontend.path}/docs/pre-dev/{feature}/
└── tasks.md           # Frontend tasks only

Note: For multi-repo, there is no central index. Each repo contains only its relevant tasks.

Task Splitting Logic

def split_tasks_by_module(all_tasks: list, topology: dict) -> dict:
    """
    Split tasks into module-specific files.

    Returns dict with keys: 'index', 'backend', 'frontend'
    """
    structure = topology.get('structure', 'single-repo')
    modules = topology.get('modules', {})
    backend_path = modules.get('backend', {}).get('path', '.')
    frontend_path = modules.get('frontend', {}).get('path', '.')

    backend_tasks = [t for t in all_tasks if t.get('target') == 'backend']
    frontend_tasks = [t for t in all_tasks if t.get('target') == 'frontend']
    shared_tasks = [t for t in all_tasks if t.get('target') == 'shared']

    if structure == 'single-repo':
        return {
            'index': {
                'path': f"docs/pre-dev/{feature}/tasks.md",
                'tasks': all_tasks
            }
        }

    if structure == 'monorepo':
        return {
            'index': {
                'path': f"docs/pre-dev/{feature}/tasks.md",
                'tasks': all_tasks
            },
            'backend': {
                'path': f"{backend_path}/docs/pre-dev/{feature}/tasks.md",
                'tasks': backend_tasks + shared_tasks
            },
            'frontend': {
                'path': f"{frontend_path}/docs/pre-dev/{feature}/tasks.md",
                'tasks': frontend_tasks + shared_tasks
            }
        }

    if structure == 'multi-repo':
        return {
            'backend': {
                'path': f"{backend_path}/docs/pre-dev/{feature}/tasks.md",
                'tasks': backend_tasks + shared_tasks
            },
            'frontend': {
                'path': f"{frontend_path}/docs/pre-dev/{feature}/tasks.md",
                'tasks': frontend_tasks + shared_tasks
            }
        }

Module-Specific Task File Header

Each module-specific tasks.md should include:

---
feature: {feature-name}
module: backend | frontend
filtered_from: docs/pre-dev/{feature}/tasks.md  # (monorepo only)
total_tasks: N
---

# {Feature Name} - {Module} Tasks

This file contains tasks filtered for the **{module}** module.

**Full task list:** {link to index if monorepo, or note "distributed" if multi-repo}

---

Validation for Multi-Module

Check	Requirement
All tasks have target:	If topology is monorepo or multi-repo
All tasks have working_directory:	If topology is monorepo or multi-repo
Target matches task content	Backend tasks have backend work, etc.
Working directory resolves correctly	Path exists or will be created

---

Task Template Structure

Output path depends on topology — see Output & After Approval for the full topology-dependent rules. The file starts with two summary sections followed by the full task details.

File Summary Sections (top of tasks.md)

MUST open with two summary tables before the individual task details.

Summary Table 1 — Technical Overview

A quick-reference table for the engineering team. The Status column is initialized by bee:pre-dev-task-breakdown and updated by bee:dev-cycle during execution.

• CANNOT set any value other than ⏸️ Pending at task creation time

## Summary

| Task | Title | Type | Hours | Confidence | Blocks | Status |
|------|-------|------|-------|------------|--------|--------|
| T-001 | Project Foundation | Foundation | 3.0 | High | All | ⏸️ Pending |
| T-002 | ... | Feature | 6.5 | Medium | T-004, T-008 | ⏸️ Pending |
| | **TOTAL** | | **85.0h** | | | |

MUST leave the Status cell of the TOTAL row empty. CANNOT apply ⏸️ Pending or any status value to the TOTAL row.

Status lifecycle (managed by bee:dev-cycle):

Value	Meaning	Set by
⏸️ Pending	Not started	bee:pre-dev-task-breakdown at task creation
🔄 Doing	Execution started (Gate 0 began)	bee:dev-cycle
✅ Done	Gate 9 approved	bee:dev-cycle
❌ Failed	Execution terminated with unresolved blocker	bee:dev-cycle

Summary Table 2 — Business Deliverables View

MUST appear immediately after Summary Table 1. A plain-language view for product managers, stakeholders, and the team to understand what value each task delivers.

## Business Deliverables

| Task | Deliverable (business view) |
|------|-----------------------------|
| T-001 | The team can develop and test locally from day one — **every contributor gets a working environment without manual setup**. |
| T-002 | **Transactions reach their destination** — messages conform to the agreed standard and counterparties accept every one sent. |
| ... | _(additional tasks omitted for brevity)_ |

Writing rules for Business Deliverables View:

Rule	Correct	Wrong
Language	Plain business language	Technical jargon (endpoints, migrations, repositories)
Perspective	What the business/user gains	What the developer implements
Voice	Active — "The product can...", "Users gain..."	Passive — "It is implemented...", "It will be created..."
Length	1-3 sentences max	Bullet lists, long paragraphs
Emphasis	Bold the core value proposition	No bold or no emphasis at all
Source	Derived from each task's Deliverable field	Invented separately

What to include:

• The capability unlocked ("The product can...", "Users can now...")
• Why it matters to the business ("Without this, no transaction reaches its destination")
• Who benefits (developer, operator, end user, regulator)
• Business consequence if missing (when relevant)

What to exclude:

• File names, class names, function names
• Architecture terms (API, REST, handler, service layer)
• Infrastructure terms (Docker, Kubernetes, PostgreSQL)
• Implementation verbs (implement, create endpoint, configure database)

Validation for Business Deliverables View:

Check	Requirement
Language	No technical jargon (no API, REST, handler, migration, repository)
Length	1-3 sentences per row; no bullet lists
Voice	Active and capability-focused; no passive constructions
Source	Each row derived from the task's Deliverable field, not invented
Formatting	Core value proposition bolded; no other inline formatting
Exclusions	No file/class names, architecture terms, infrastructure terms, or implementation verbs

---

Per-Task Template

Each task includes:

Section	Content
Header	T-[XXX]: [Task Title - What It Delivers]
Target	backend | frontend | shared (if multi-module)
Working Directory	Path from topology config (if multi-module)
Agent	Recommended agent: bee:backend-engineer-, bee:frontend--engineer-*, etc.
Deliverable	One sentence: what working software ships
Scope	Includes (specific capabilities), Excludes (future tasks with IDs)
Success Criteria	Testable items: Functional, Technical, Operational, Quality
User/Technical Value	What users can do; what this enables
Technical Components	From TRD, From Dependencies
Dependencies	Blocks (T-AAA), Requires (T-BBB), Optional (T-CCC)
Effort Estimate	AI Estimate: X AI-agent-hours, Confidence: [High/Medium/Low], Estimation Method: [Agent Name], Team type
Risks	Per risk: Impact, Probability, Mitigation, Fallback
Testing Strategy	Unit, Integration, E2E, Performance, Security
Definition of Done	Code reviewed, tests passing, docs updated, security clean, performance met, deployed to staging, PO acceptance, monitoring configured

AI Estimation Fallback Procedure

When AI estimation fails or is unavailable:

Failure Conditions

AI estimation is considered failed when:

• ❌ Agent unreachable (API timeout >30s)
• ❌ API errors (rate limit exceeded, 500 errors)
• ❌ Service outage (Claude API down)
• ❌ Invalid response (no estimate returned)
• ❌ Scope too ambiguous (agent returns Low confidence with no estimate)

Escalation Path

1. Immediate (0-5 min): Retry once with exponential backoff
2. Short-term (5-15 min): Switch to backup agent if available
3. Medium-term (15-30 min): Notify PM team lead via Slack
4. Long-term (30+ min): Use manual estimation override (see below)

Manual Estimation Override

Who can approve: PM Team Lead or designated backup

Required evidence for override:

• Similar historical task with actual time data
• Expert estimation from technical lead (with rationale)
• Decomposed scope with per-component estimates
• Documentation of estimation method used

How to record:

**Effort Estimate:**
- AI Estimate: [FAILED - API unavailable]
- Manual Override: X hours (approved by: [Name], date: YYYY-MM-DD)
- Estimation Method: Historical comparison with Task T-XXX
- Confidence: Medium (manual estimation, subject to higher variance)
- Evidence: [Link to similar task or rationale document]

Temporary Capacity Adjustments

When manual estimation is used:

• Apply 1.3x multiplier to manual estimates (30% buffer for estimation uncertainty)
• Mark task with "Estimation Pending" flag in roadmap
• Schedule re-estimation when AI service restored
• Track variance post-completion to calibrate manual estimates

Example:

Manual estimate: 6 hours
Adjusted estimate: 6h × 1.3 = 7.8 hours
Confidence: Medium → Low (due to estimation method)
Re-estimation scheduled: [Date when AI available]

Confidence Flag Requirements

Align with rationalization table:

• ✅ Manual estimates MUST be marked with confidence level
• ✅ Include rationale for estimation method
• ✅ Link to historical data or expert judgment
• ✅ Document who approved and when
• ❌ NEVER mark manual estimates as "High" confidence
• ❌ NEVER skip variance tracking for manual estimates

Common Violations

Violation	Wrong	Correct
Technical-Only Tasks	"Setup PostgreSQL Database" with install/configure steps	"User Data Persistence Foundation" with deliverable (working DB layer <100ms), user value (enables T-002/T-003), success criteria (users table, pooling, migrations)
Oversized Tasks	"Complete User Management System" (6 weeks) with all auth features combined	Split into: T-005 Basic Auth (L), T-006 Password Mgmt (M), T-007 2FA (M), T-008 Permissions (L)
Vague Success Criteria	"Feature works, Tests pass, Code reviewed"	Functional (upload 100MB, formats), Technical (<2s response), Operational (99.5% success rate), Quality (90% coverage)

Delivery Sequencing

Optimize task order by sprint/phase with goals, critical path identification, and parallel work opportunities.

Confidence Scoring

Factor	Points	Criteria
Task Decomposition	0-30	All appropriately sized: 30, Most well-scoped: 20, Too large/vague: 10
Value Clarity	0-25	Every task delivers working software: 25, Most clear: 15, Unclear: 5
Dependency Mapping	0-25	All documented: 25, Most clear: 15, Ambiguous: 5
Estimation Quality	0-20	Based on past work: 20, Educated guesses: 12, Speculation: 5

Action: 80+ autonomous | 50-79 present options | <50 ask about velocity

Output & After Approval

Output to (depends on topology.structure):

Structure	Files Generated
single-repo	docs/pre-dev/{feature}/tasks.md
monorepo	Index + {backend.path}/docs/pre-dev/{feature}/tasks.md + {frontend.path}/docs/pre-dev/{feature}/tasks.md
multi-repo	{backend.path}/docs/pre-dev/{feature}/tasks.md + {frontend.path}/docs/pre-dev/{feature}/tasks.md

1. ✅ Tasks become sprint backlog
2. 🎯 Use as input for subtasks (bee:pre-dev-subtask-creation)
3. 📊 Track progress per task (not per subtask)
4. 🚫 No implementation yet - that's in subtasks

The Bottom Line

If you created tasks that don't deliver working software, rewrite them.

Tasks are not technical activities. Tasks are working increments.

"Setup database" is not a task. "User data persists correctly" is a task. "Implement OAuth" is not a task. "Users can log in with Google" is a task. "Write tests" is not a task. Tests are part of Definition of Done for other tasks.

Every task must answer: "What working software can I demo to users?"

If you can't demo it, it's not a task. It's subtask implementation detail.

Deliver value. Ship working software. Make tasks demoable.

---

Standards Loading (MANDATORY)

This skill is a task decomposition skill and does NOT require WebFetch of language-specific standards.

Purpose: Task Breakdown defines WHAT value increments to deliver, not HOW to implement them. Language-specific standards apply during subtask creation and implementation.

However, MUST load PRD (Gate 1), TRD (Gate 3), and research.md to ensure tasks align with requirements and architecture.

---

Blocker Criteria - STOP and Report

Condition	Action	Severity
PRD (Gate 1) not validated	STOP and complete Gate 1 first	CRITICAL
TRD (Gate 3) not validated	STOP and complete Gate 3 first	CRITICAL
Task exceeds 2 weeks (XL size)	STOP and break down further	HIGH
Task has no measurable success criteria	STOP and define testable criteria	HIGH
Task has no user or technical value	STOP and redefine as value delivery	HIGH
AI estimation failed	Follow fallback procedure in skill	MEDIUM
Dependencies form circular loop	STOP and resolve dependency cycle	HIGH

---

Cannot Be Overridden

These requirements are NON-NEGOTIABLE:

• MUST NOT create tasks exceeding 2 weeks (break them down)
• MUST include measurable success criteria for every task
• MUST include user value OR technical value for every task
• MUST include task dependencies (blocks, requires, optional)
• MUST run AI estimation for all tasks (no manual guesses)
• MUST include Definition of Done checklist for every task
• MUST include Business Deliverables View in summary section
• MUST initialize Status column as ⏸️ Pending for every task row
• CANNOT set Status to any value other than ⏸️ Pending at task creation
• CANNOT use technical-only tasks without value connection

---

Severity Calibration

Severity	Definition	Example
CRITICAL	Cannot create valid tasks	PRD/TRD not validated, no requirements to decompose
HIGH	Task violates sizing or value rules	XL task, no success criteria, no value defined
MEDIUM	Task incomplete but usable	Missing one dependency, unclear testing strategy
LOW	Minor documentation gaps	Definition of Done could be more detailed

---

Pressure Resistance

User Says	Your Response
"This 3-week task is fine"	"Cannot accept 3-week tasks. Tasks >2 weeks hide complexity. I'll break it into smaller deliverables."
"Setup tasks don't need user value"	"Cannot create valueless tasks. Setup ENABLES value. I'll define what this setup enables."
"Success criteria are obvious"	"Cannot assume obvious criteria. Obvious to you ≠ testable. I'll document explicit, measurable criteria."
"Skip AI estimation, use story points"	"Cannot skip AI estimation. Story points are abstract; AI hours are concrete. I'll run AI analysis."
"We can figure out dependencies later"	"Cannot defer dependencies. Later is too late. I'll map dependencies now."

---

Anti-Rationalization

Rationalization	Why It's WRONG	Required Action
"This 3-week task is fine"	Tasks >2 weeks hide complexity	Break down into ≤2 week tasks
"Setup tasks don't need value"	Setup enables value. Define what it enables	Connect to user/technical value
"Success criteria are obvious"	Obvious to you ≠ testable. Document explicitly	Define measurable criteria
"Dependencies will be clear later"	Later is too late. Map them now	Document all dependencies
"Skip AI estimation, use story points"	Story points are abstract; AI hours are concrete	Run AI estimation
"Technical tasks can skip user value"	Even infrastructure enables users. Define connection	Connect to user impact
"Testing strategy can be decided during"	Testing affects design. Plan it upfront	Define testing strategy now

---

When This Skill Is Not Needed

• PRD or TRD not validated (complete earlier gates first)
• Tasks already exist and are validated
• Trivial change that can be implemented directly
• Bug fix with known solution (no decomposition needed)
• Documentation-only updates
• Configuration changes with no implementation work