Decomposition: Least-to-Most & Subproblems
Break a hard task into ordered subquestions, solve them one by one, and feed each answer forward — reasoning as a pipeline, not a leap.
Watch the monolithic prompt fudge it
Give a model a compound question in one gulp and it tends to answer the vivid parts, skim the rest, and blend everything into confident mush. Run this and grade it honestly: which required considerations got real numbers, and which got hand-waves?
Our SaaS charges $9/month and has 2,000 subscribers, 4% monthly churn, and ~300 new signups a month from ads costing $3,000/month. Should we raise the price to $12? Consider revenue, likely churn impact, how the change affects payback on ad spend, and what we should test first. Give a recommendation.VerifyYou get a fluent answer — and on close reading, at least one required consideration was handled with a hand-wave instead of arithmetic.Ask for the decomposition, not the answer
Least-to-most prompting starts by having the model produce the subquestion ladder — ordered from what's answerable now to what depends on everything else. Forbid answering: you want the plan alone, so you can inspect and fix it before any reasoning builds on it.
I need to decide whether to raise our SaaS price from $9 to $12/month. Do NOT answer the question yet. First, list the subquestions I'd need to answer, ordered so each one depends only on the ones before it. Start from simple factual calculations and build toward the final recommendation. Number them.VerifyYou get an ordered ladder — current revenue math before churn scenarios before break-even before recommendation — with no answers attached.Solve the first rungs in isolation
Now work the ladder from the bottom, one subquestion per prompt. Early rungs are usually pure arithmetic the model handles near-perfectly in isolation — exactly the parts a monolithic answer fudges while narrating.
Context: 2,000 subscribers at $9/month, 4% monthly churn, 300 signups/month, $3,000/month ad spend. Answer ONLY subquestion 1: What is current MRR, and what would MRR be at $12 with the subscriber count unchanged? Show the arithmetic. Do not discuss churn or the recommendation yet.VerifyClean, checkable numbers ($18,000 and $24,000) with no premature opinions leaking in.Feed answers forward
Each later rung receives the established results pasted in as settled fact — that's the chaining in least-to-most. The model reasons about one new thing at a time while standing on ground you've already verified.
Established so far: - Current MRR: $18,000. At $12 with no churn change: $24,000 (+$6,000). - Break-even: the raise still wins if we retain at least 1,500 subscribers (75%). Answer ONLY this subquestion: For a ~30% price increase on a $9 product, what churn outcomes are plausible (best / expected / worst case), and what does MRR look like in each? Use the established numbers. Do not give the final recommendation yet.VerifyThe model reasons only about churn scenarios, anchored to your verified break-even — no re-deriving, no drift.Synthesize — and decide when the pattern is worth it
The final prompt assembles every established answer and asks only for the judgment. Use the full multi-call pattern for compound, high-stakes, or chronically mushy tasks; for tasks whose structure you already know, hardcode the subquestion sequence into a single prompt and keep the discipline without the round trips.
Established: 1. MRR: $18,000 today; $24,000 at $12 if churn holds. 2. Break-even retention: 75% of current subscribers. 3. Churn scenarios: best +$5,400/mo, expected +$3,600/mo, worst -$900/mo. Using only these established results: give a recommendation, name the single riskiest assumption, and propose the cheapest test to run before committing (e.g. new-signups-only pricing, grandfathering existing users). Three short paragraphs.VerifyThe recommendation cites the established numbers — and you could audit every figure it rests on, because you already did.