Reasoning Patterns Skill

Goal

Choose the right amount and shape of reasoning for a task, then apply it without overspending. Reasoning techniques trade extra tokens and latency for accuracy on hard problems — they are powerful on multi-step logic and actively harmful on simple lookups, where they add cost and can talk the model out of a correct first answer. The objective is to match the technique to the problem's structure (linear, branching, decomposable, abstract, environment-grounded), spend the minimum compute that clears the quality bar, and always pair self-evaluation with an external anchor.

Steps

1. Classify the problem before choosing a technique
   • Simple/lookup/well-specified → answer directly; reasoning adds latency and risk. Stop here.
   • Linear multi-step (arithmetic, logic, analysis) → chain-of-thought.
   • Noisy single trace, aggregable answer → self-consistency.
   • Search/backtracking structure (puzzles, planning, proofs) → tree-of-thoughts.
   • Decomposable into ordered easy sub-problems → least-to-most.
   • Model rushes to a specific wrong answer → step-back to the governing principle first.
   • Needs external information or to act → ReAct (defer to agentic-patterns.skill).