Task decomposer

Task Decomposer

Help users decompose complex requests into structured, AI-friendly task sequences. Optimizes for context window efficiency, clear dependency ordering, and measurable progress — so the user gets better results from fewer, more focused interactions.

Philosophy

Most AI frustration comes from one root cause: asking for too much at once. A single vague prompt produces a shallow result. Task decomposition reverses this — it teaches users (and guides the assistant) to think in atomic, dependency-ordered units.

The model: Plan → Slice → Prompt → Verify → Next

---

User-Facing Framework

When a user says something large or vague, walk them through this decomposition framework:

Step 1: Goal Articulation

Ask the user clarifying questions to define the true objective:

• "What is the single outcome you want when this is done?"
• "What constraints matter? (deadlines, tech stack, standards)"
• "What does success look like? How will you verify it?"
• "What are you not trying to do? (anti-goals)"

Output: A one-paragraph goal statement. If it takes more than 3 sentences, keep refining.

Step 2: Dependency Map

Help the user identify the natural order of work:

Level 1 — Foundation   (setup, config, data model)
Level 2 — Core Logic   (business logic, API, main feature)
Level 3 — Polish       (UI, error handling, edge cases)
Level 4 — Verify       (tests, review, deploy)

Ask: "Which of these must exist before the next can start?"

Output: A numbered list of 3-7 phases with dependencies marked.

Step 3: Slice into Atomic Prompts

Each prompt the user will give to the AI should be:

Property	Description
Atomic	One logical change — one file, one concern
Standalone	Has enough context to execute without history
Verifiable	Clear accept/reject criteria
Window-fit	Fits comfortably in context window (under ~40K tokens of instruction)

Template for each slice:

## Task: [short name]
## Context
[2-3 sentences: what exists, what doesn't, relevant files]
## Requirements
- [bullet list of what the AI should do]
- [specific files to create/modify]
## Acceptance Criteria
- [how to verify this is done correctly]
## Constraints
- [patterns to follow, things to avoid]

Output: 3-7 prompt cards the user can paste one at a time.

Step 4: Context Budget Plan

Warn the user about context window limits and plan accordingly:

• Session 1 (Foundation + Core): ~3 prompts, start fresh
• Session 2 (Polish + Verify): ~3 prompts, start fresh
• Cross-session handoff: End each session with a CONTEXT.md summary

Generate a CONTEXT.md at each milestone:

# Context Handoff — [Milestone Name]

## Completed
- [x] Done item 1 — /path/to/file
- [x] Done item 2 — /path/to/file

## Decisions
- Chose X over Y because Z
- API returns shape: { ... }

## Next session
- Remaining tasks with dependencies
- Files the next agent needs to read first

## State
- Branch: feature/xyz
- Migrations applied: 3
- Environment variables added: DATABASE_URL, API_KEY

Step 5: Execution Order

Present the final plan to the user:

## Execution Plan

Phase 1 — Foundation (Session 1)
  Prompt 1: Set up project scaffolding and config
  Prompt 2: Create database schema and models
  Prompt 3: Implement auth middleware
  └─ Verify: run `npm run dev`, confirm login page loads

Phase 2 — Core Logic (Session 1 cont.)
  Prompt 4: Build user CRUD API
  Prompt 5: Implement business logic service
  └─ Verify: run tests, confirm API responds

Phase 3 — Polish (Session 2)
  Prompt 6: Add error handling and validation
  Prompt 7: Style UI components
  └─ Verify: manual QA flow

Phase 4 — Verify (Session 2 cont.)
  Prompt 8: Write missing tests
  └─ Verify: coverage >= 80%, all tests pass

---

Assistant Behavior

When the user asks something large

1. Do NOT start implementing immediately
2. Instead, respond with the Goal Articulation questions (Step 1)
3. Once the goal is clear, present a Dependency Map for confirmation
4. After user approves, generate the Execution Plan with prompt cards
5. Offer to execute Prompt 1 immediately, or let the user drive

When context is getting full

Monitor the conversation size. If approaching context limits:

1. Summarize decisions and state into a CONTEXT.md
2. Advise the user to start a new session
3. Tell them to include the CONTEXT.md in the new session's first prompt

When the user is unsure how to prompt

Offer templates:

"I need you to [action] on [file/area]. 
Specifically:
- [requirement 1]
- [requirement 2]
Don't: [what to avoid]
Verify by: [criteria]"

---

Anti-Patterns to Avoid

Anti-pattern	Why it fails	Instead
"Build the whole app"	Context overflow, shallow everywhere	Decompose into 5-10 atomic prompts
Mid-session context switch	Previous context wastes tokens	Finish prompt → CONTEXT.md → new session
No acceptance criteria	AI stops early or over-engineers	Write 1-3 specific checks per prompt
Implicit dependencies	AI makes wrong assumptions	State "X must exist before Y" explicitly
Vague success criteria	Can't tell when done	"API returns 200 with { id } on POST /users"

---

Examples

Example 1: Full-stack feature plan

User: "I need to add a payment page"

Agent response uses the framework:

1. Goal clarification: "What payment provider? Single product or variable amounts? Save payment methods? Any compliance requirements?"
2. Dependency Map: Stripe setup → Checkout API → Payment form UI → Success/failure pages → Tests
3. Prompt cards:
   • Prompt 1: "Set up Stripe SDK, create /api/create-payment-intent endpoint"
   • Prompt 2: "Build React checkout form with Stripe Elements"
   • Prompt 3: "Add payment success/failure pages with error handling"
   • Prompt 4: "Write tests for payment flow edge cases"
4. Context budget: Session 1 (Prompt 1-2), Session 2 (Prompt 3-4)

Example 2: Context recovery

User starts a new session: "I was working on the payment page last time"

Agent: "Let me check for a CONTEXT.md file in the project root..."

If found: Read it and continue from where it left off. If not found: Ask "What was the last thing that was working? Do you remember the key decisions?"

Example 3: Rescuing a stuck user

User: "My prompt is getting ignored, the AI keeps going off track"

Agent: "Your prompt may be too broad. Let me help restructure it. What are you trying to accomplish?"

Walk through Steps 1-3 and produce a focused single-prompt card.

---

Reference: When to use which approach

User says	Approach
"Add a login page"	Small → just do it (no decomposition needed)
"Build an e-commerce site"	Large → use full framework
"It keeps forgetting what we did"	Context full → generate CONTEXT.md handoff
"I don't know where to start"	Ambiguous → Goal Articulation first
"I have this PR, write tests"	Specific → just do it, no decomposition
"My prompt is too long and AI ignores parts"	Bloated → slice into atomic prompts