Good chains make the model expose assumptions, compare alternatives and show where confidence is weak before it gives the final recommendation.
Reasoning patterns
Decision-grade chains
Failure drop
Against naive prompts
Auditability
Traceable structure
Best context
Long-form tasks
| Pattern | Best for | Guardrail | Failure mode |
|---|---|---|---|
| Plan -> Verify -> Answer | Architecture reviews | Force explicit assumptions | Shallow verification |
| Hypothesis tournament | Debugging | Competing root causes | Premature convergence |
| Constraint ladder | Migration plans | Lock non-negotiables first | Scope drift |
| Decision memo | Tradeoff analysis | State alternatives rejected | One-sided recommendation |
Reasoning block
For architecture, vendor or rollout choices where rejected options must stay visible.
Write a decision memo with context, constraints, options considered, rejected alternatives, chosen option, risk register and next validation step.
Reasoning block
For debugging or incident review when premature convergence is dangerous.
List the three most likely root causes, the evidence for each one, the evidence against each one and the fastest discriminating check before recommending a fix.
Reasoning block
For migrations and plans where the non-negotiables must be locked first.
Before proposing a plan, extract non-negotiable constraints, rank them, note open assumptions and then build the plan around them.
Best entry point when the team needs a stack recommendation, migration plan and hard constraints in one output.
Use when reasoning must end in code changes, tests and operational steps.
Reasoning gets better when context retention, retrieval and state transitions are designed upfront.