reasoning-models

Reasoning Models

Purpose: Get reliable answers from reasoning models without overspending or stalling on tasks that do not need them.

---

When to Use This Skill

• Choosing between a reasoning model (Claude Opus 5.5, GPT-6 Astra, DeepSeek R1, Gemini Thinking) and a fast model
• Setting effort (low / medium / high, plus xhigh / max where supported)
• Designing prompts for reasoning models (different from chat models)
• Combining reasoners (planner) with fast models (executor)
• Diagnosing high cost or latency on reasoning calls

When NOT to Use a Reasoning Model

• Routine extraction, summarization, classification -- a fast model is cheaper and faster
• Strict latency budget (<2s end-to-end) -- reasoning models add seconds
• Tool-use-heavy loops where each turn is short -- fast model + scaffolding is usually better

---

Decision Tree

Is the task hard? (multi-step reasoning, math, planning, code refactor across files,
                   ambiguous spec, agent strategy)
+- No  -> Fast model or the lowest effort that passes evals (claude-haiku-4.5, gpt-6-luna)
+- Yes -> Reasoning model
        +- Need fast feedback loop?  -> Effort = low / medium
        +- Quality > latency?         -> Effort = high; xhigh / max only after a measured gain
        +- Multi-turn tool use?       -> Often: fast model with strong scaffold beats a reasoner alone

---

Prompt Patterns That Differ for Reasoners

Pattern	Chat / Fast Model	Reasoning Model
Chain-of-thought ("think step by step")	Helpful	Counterproductive; Opus 5.5 may refuse (reasoning_extraction)
Few-shot examples	Often improves	Use sparingly; can over-anchor
Long detailed system prompts	Often needed	Often shorter prompts work better
Strict output format	Add explicit schema	Use Structured Outputs / response_format
Self-critique loops	Sometimes helps	Built-in; do not duplicate

Rule of thumb: give reasoning models the goal and the constraints, not the procedure.