Ask AI to do ten things at once, and it does all ten poorly. Ask it to do one thing, then the next, then the next, and each one works. The difference isn’t magic. It’s attention.
Why Single Prompts Fail
You ask AI to write a business plan. It produces something that looks like a business plan. Executive summary, market analysis, financial projections, all there. But the market analysis contradicts the financial assumptions. The executive summary doesn’t match the details. The whole thing feels thin.
This happens because AI attention is finite. Complex prompts spread that attention across many tasks simultaneously. Each task gets a fraction of the focus. Quality drops everywhere.
The fix is obvious once you see it: don’t ask for everything at once. Ask for pieces, one at a time.
The Decomposition Principle
Instead of: “Write a complete business plan for my startup.”
Try:
- “Create an outline for a business plan in [domain]. Just the structure, no content yet.”
- “Now write the market analysis section based on this outline.”
- “Now the competitive analysis.”
- “Now financial projections.”
- “Review all sections together. Flag any contradictions or inconsistencies.”
Each prompt is bounded. AI focuses on one thing. Quality improves on each piece. Errors in one section don’t contaminate others. You catch problems early, fix them before they propagate.
Google Research formalized this in 2022 as “Least-to-Most Prompting” (Zhou et al.). Their finding: breaking complex problems into sequential subproblems produced better results than direct attacks on the whole, particularly on reasoning tasks.
A caveat: This research came from the GPT-3 era. Models have improved substantially since then. GPT-4, Claude 3.5, and later versions handle complex instructions far better than their predecessors. The decomposition advantage has shrunk. But it hasn’t disappeared, especially for long outputs, multi-part documents, or anything where you need to verify as you go.
When Decomposition Helps
Not every task needs breaking down. The overhead isn’t free.
Decompose when:
The output has multiple sections serving different purposes. A report with analysis, recommendations, and action items. A document with research, synthesis, and conclusions.
Later parts depend on earlier parts being correct. If the outline is wrong, every section built on it is wrong. Better to verify the outline first.
Length exceeds a page or two. AI consistency degrades over length. The longer the output, the more likely decomposition helps.
You need to verify as you go. Simple outputs are easier to check. “Is this outline right?” is answerable. “Is this entire 20-page document right?” is overwhelming.
Don’t decompose when:
The task is genuinely atomic. One question, one answer, done.
Speed matters more than perfection. Decomposition takes longer. If you need something fast and good-enough, single prompt may be better.
You’re exploring, not producing. Early iterations don’t need the overhead. Decompose when you’re building something final.
The Hidden Cost
Decomposition creates a new problem: the pieces might not fit together.
Section 2 uses different terminology than Section 1. The tone shifts halfway through. Conclusions in Part 4 contradict assumptions in Part 1. Each piece is good; the whole is incoherent.
You have to manage consistency actively:
Pass context forward. Each new prompt includes relevant output from previous steps. “Based on this outline: [outline], write section 2. Maintain consistency with the approach established.”
Lock terminology early. “Use ‘customer acquisition cost’ throughout, not ‘CAC’ or ‘acquisition expense.'”
Review integration explicitly. After generating all pieces, run a final pass: “Review these sections together. Identify any contradictions, tone shifts, or terminology inconsistencies.”
This takes time. Factor it in. Total time equals generation time plus review time plus integration time. Decomposition isn’t faster. It’s better.
The Real Skill
Here’s what most articles about decomposition miss: knowing how to break down a task is itself a skill.
Bad decomposition produces bad results. If you break a business plan into “good parts” and “boring parts,” you haven’t helped. If your pieces are still too complex, you haven’t decomposed enough. If they’re too granular, you’ve created overhead without benefit.
Good decomposition requires understanding the task’s structure. What depends on what? What can be done independently? Where are the natural boundaries?
Three questions to ask:
- What’s the first thing that has to be right before anything else can work? Start there.
- What pieces are independent versus sequential? Independent pieces can be done in any order. Sequential pieces have dependencies.
- Where would an error in this piece corrupt everything downstream? Those are your verification checkpoints.
If you can’t answer these questions, you don’t understand the task well enough to decompose it. That’s useful information too.
A Practical Example
Here’s a real comparison I ran:
Single prompt: “Write a market analysis report on electric vehicles including market size, competitive landscape, and trends.”
Result: 1200 words. Market size section referenced “major players” before the competitive section introduced them. Trends section contradicted a growth figure from market size. Tone shifted from analytical to promotional halfway through.
Decomposed approach (8 prompts):
- “What are the key sections a thorough EV market analysis should include?” (Structure first)
- “For the market size section: current global figures and growth projections. Bullet points.” (Data gathering)
- “For competitive landscape: major players, market positions, recent moves.” (Data gathering)
- “For trends: technology, regulatory, consumer behavior shifts.” (Data gathering)
- “Draft the market size section as narrative, based on the bullets.” (Synthesis)
- “Draft competitive landscape section.” (Synthesis)
- “Draft trends section.” (Synthesis)
- “Review all sections. Are there contradictions? Does the analysis flow? Are conclusions supported by the data presented?” (Integration)
Result: 1400 words. Each section internally consistent. Integration review caught two terminology mismatches, fixed before final output. No contradictions.
Time difference: Single prompt took 2 minutes. Decomposed took 15 minutes including review.
The decomposed version wasn’t dramatically better in every way. It was verifiable. I could check each piece as it emerged. The single-prompt version required reading the whole thing to find problems, then figuring out how to fix them without creating new ones.
Note the difference from bad decomposition like “write the first half / write the second half.” That splits by length, not function. Useless. Good decomposition follows the task’s natural structure.
When to Stop Decomposing
You can over-decompose. Every sentence becomes a prompt. The overhead drowns the benefit.
Signs you’ve gone too far:
You’re spending more time managing prompts than reviewing output.
The pieces are so small they lack context. AI needs enough information to produce something meaningful.
You’re decomposing out of anxiety rather than necessity. Not every task is complex. Sometimes single prompts work fine.
The goal isn’t maximum decomposition. It’s appropriate decomposition. Complex, high-stakes, multi-part work benefits most. Simple, fast, exploratory work usually doesn’t need it.
The Bottom Line
AI attention is finite. Complex prompts dilute it. Simple, sequential prompts concentrate it.
Decompose tasks that have multiple parts, sequential dependencies, or length beyond a page or two. Don’t decompose tasks that are genuinely simple or where speed beats quality.
The skill isn’t just prompting. It’s knowing how to break work into pieces that AI can handle well and you can verify easily. That skill transfers beyond AI. It’s how complex work gets done, with or without tools.
Good decomposition means understanding your task’s structure. If you can’t break it down, you might not understand it well enough yet. That’s worth knowing before you start.