The Prompt Lab — Iterative Decomposition
Learn the iterative decomposition prompting technique with concrete before/after examples.
Iterative Decomposition
The Technique
Iterative Decomposition means breaking a complex task into an explicit sequence of smaller sub-tasks, asking the model to complete each one before moving to the next — either in separate prompts or by instructing it to show its work stage by stage. This works because frontier models handle bounded, well-defined tasks more reliably than open-ended requests that quietly require ten decisions to happen simultaneously and correctly.
The Naive Prompt
Write a cold email to a CFO at a mid-size SaaS company selling our expense
management software, Finlo. Make it persuasive and professional.Why It Falls Short
The model has to silently guess the CFO’s pain points, decide what “persuasive” means for this persona, invent a value proposition, choose a tone, and structure the email — all at once, with no visibility into any of those choices. The output is usually competent but generic: confident-sounding sentences that could apply to almost any B2B product. You get a draft, not a reasoned draft you can actually improve.
The Improved Prompt
I'm writing a cold email to a CFO at a mid-size SaaS company (200–800 employees)
to sell Finlo, our expense management software. Work through this in stages and
show each stage before moving on:
Stage 1 — CFO Pain Points: List 4–5 specific problems a CFO at this company
size likely has with expense management right now.
Stage 2 — Hook Selection: From those pain points, choose the single strongest
hook for a cold email and explain why it beats the others.
Stage 3 — Value Proposition: Write one sentence connecting that pain point to
what Finlo solves, using concrete language (time, money, or risk).
Stage 4 — Draft the Email: Now write the cold email using the hook and value
prop from stages 2–3. Max 120 words. One CTA at the end.Why It Works
Each stage produces a visible, inspectable output — so if the model chooses a weak hook in Stage 2, you catch it before it contaminates the final draft. The staged structure also forces the model to commit to a specific reasoning path rather than averaging across all plausible approaches, which is exactly what turns generic prose into something targeted. You end up with both a better email and a working document you can tweak at any stage.
When to Use This
- High-stakes single outputs — onboarding sequences, investor memos, legal summaries — where a plausible-sounding but poorly-reasoned draft is actively dangerous.
- Tasks with hidden decisions — any prompt where the model must make several non-obvious judgment calls before the “real” work begins. Decomposition makes those calls auditable.
- Long-context models with drift — when working in extended sessions with GPT-5.5 or Claude Opus 4.7, staging prevents the model from losing the thread of an earlier decision by the time it reaches the final output.