Self-consistency improves on chain-of-thought by generating multiple independent reasoning paths and selecting the most consistent answer — instead of trusting a single, potentially flawed line of logic.
Especially effective for arithmetic, logic puzzles, and judgment calls where one reasoning path can confidently go wrong. The intuition: if 4 out of 5 independent attempts converge on the same answer, it's far more trustworthy than any single attempt.
02 Weak vs. Strong
EX 01Pricing decision with 3 independent paths
We're pricing Vitae's Pro tier: $19 vs $29/month. Context: resume analytics SaaS, target = mid-career tech professionals, competitors at $15 (Teal) and $25 (Rezi), our differentiator is recruiter-view analytics nobody else has.
Analyze this 3 times, completely independently — do not let one analysis influence another:
ANALYSIS 1 — Value-based lens: what is the analytics feature worth to someone in an active job search? Work from outcomes (interview rate, time-to-offer).
ANALYSIS 2 — Competitive positioning lens: work only from the competitor map and price-quality signaling.
ANALYSIS 3 — Unit economics lens: work only from conversion and churn math. Assume price elasticity: each $10 increase costs ~15% of conversions.
Then: state each analysis's conclusion in one line. If they agree, that's the answer. If they diverge, identify exactly which assumption drives the disagreement — that's what we need to test.
→ Why it works
Three genuinely different analytical frames, run independently. Convergence = confidence. Divergence pinpoints the assumption to validate — far more useful than one confident-sounding recommendation.
EX 02Catching a calculation error
Calculate the break-even point for this scenario:
- Fixed costs: $4,200/month (infra + tools)
- Price: $24/month per user
- Variable cost: $3.10/user/month (payment fees + compute)
- Current users: 140
Solve this 3 separate times using different methods:
Method 1: Contribution margin formula
Method 2: Build a month-by-month table from 0 users until profit ≥ 0
Method 3: Algebraic: solve 24x − 3.10x − 4200 = 0 step by step
Show all three calculations fully. If all three agree, state the answer with confidence. If any disagree, find the arithmetic error before answering.
→ Why it works
Arithmetic is exactly where a single chain-of-thought can confidently slip. Three methods cross-check each other — an error in one is caught by the other two. 'Find the error before answering' makes disagreement productive.
03 Key Points
01Run the same reasoning task 3–5 times independently, compare conclusions
02Converging answers = high confidence. Diverging = the task is genuinely ambiguous
03In one prompt: 'Solve this 3 ways independently, then compare and pick'
04Divergence is diagnostic — it shows WHERE the reasoning is fragile
05Cost-aware version: cheap model × 5 paths often beats expensive model × 1
04 Model-Specific Notes
Claude maintains genuine independence between analyses when instructed. For critical decisions, run paths in separate conversations for true independence.
05 For Your Role
For any important answer: 'Now solve this again from scratch a different way. Same answer?'