Decomposition: Break the Big Task into Small Steps

Some prompts ask for too much at once. "Plan a whole product launch." "Write a full course." "Migrate this entire app." The model will produce something — but it'll be vague, unevenly developed, and hard to act on. Decomposition is the pattern where you ask the model to break the task into smaller, individually solvable pieces before doing any of them.

A worked example

A planning prompt with explicit shape:

I want to launch a paid newsletter. Break this into a 6-step plan, each step
one sentence, each step assignable to a single afternoon of work. Do not
write the newsletter content itself.

Captured output:

# 6-Step Plan to Launch a Paid Newsletter

**Step 1:** Research 3-5 successful paid newsletters in your niche, noting their pricing, frequency, topics, and value propositions to identify what works.

**Step 2:** Choose your newsletter platform (Substack, beehiiv, Ghost, etc.), create your account, and configure basic settings including name, description, and pricing tier.

**Step 3:** Design a simple landing page with your value proposition, sample topics you'll cover, pricing, and sample topics you'll cover, then write your welcome email sequence.

**Step 4:** Create promotional graphics and write 5-7 social media posts announcing your newsletter across your existing platforms (LinkedIn, Twitter, etc.).

**Step 5:** Reach out personally to 20-30 contacts via email or DM who might be interested, offering them a founding member discount or free trial.

**Step 6:** Set up your payment processing, test the entire subscriber journey from sign-up to payment to email delivery, and schedule your official launch date announcement.

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

Notice the shape:

• Six steps, each one sentence. Hard constraint, honoured.
• Each step is afternoon-sized. Not "build the whole funnel" — specific, scoped, actionable.
• The model resisted scope creep. "Do not write the newsletter content itself" was honoured — no sample copy snuck in.

Why decomposition is so useful

Three reasons:

1. You can hand off individual steps. Step 1 might go to research, step 4 to design. The decomposition turns one big problem into a delegation tree.
2. You can iterate one step at a time. Don't like step 5? Re-prompt just that step. Steps 1–4 stay stable.
3. You catch missing assumptions early. The model might surface a step you hadn't thought of (in this case, step 5 about reaching out to founding members). Better to see that now than after launch.

How to design a decomposition prompt

The prompt above hides four important constraints. Steal them:

Constraint	Effect
"Break this into a 6-step plan"	Sets a fixed number — prevents 3-step underspecified plans or 20-step overwhelm.
"Each step one sentence"	Forces concision. Stops the model from writing a sub-essay per step.
"Each step assignable to a single afternoon of work"	Sets the granularity of each step.
"Do not write the newsletter content itself"	Prevents scope creep into the next-level task.

If you write decomposition prompts without these, you get hand-wavy output. With them, you get a concrete plan.

When to chain it with other patterns

Decomposition pairs beautifully with later patterns. A common chain:

1. Decompose the big task into 6 steps.
2. For each step, use few-shot or chain-of-thought to actually do that step.
3. After each step, run self-critique to verify the result.

That's not three separate prompts you write from scratch — it's the same building blocks composed. Module 5 will show this in action when we repair a bad LinkedIn rewrite.

A small habit

Anytime you catch yourself writing a prompt that asks for "everything" — a full launch plan, a complete migration, a whole course outline — pause and decompose first. The 30 seconds you spend asking for the plan saves you the 30 minutes you'd spend re-prompting because the first attempt was a uneven blob.

Module 4 begins: stop talking about patterns and start using them on real day-to-day work. :::