Prompting Reasoning Models

Advanced

Stop writing the thinking — start writing the spec

01 The Concept

Reasoning models (Claude's extended thinking, OpenAI's o-series, Gemini's thinking modes) run an internal deliberation phase before answering. This flips several classic prompt habits: 'think step by step' is redundant — the model already will — and micromanaging how to reason ('first list X, then compare Y…') can actually hurt, constraining a search process that's better left free.

What reasoning models want is a clean problem spec: all the facts, the actual constraints, and an unambiguous definition of success. They will faithfully over-deliberate a mis-specified problem. Your leverage moves to three places: choosing when reasoning is worth it (multi-step logic, math, tricky tradeoffs — not lookup or formatting), setting the effort/budget dial where the API offers one, and writing verification hooks — 'state the answer, then list the assumptions that would break it'.

02 Weak vs. Strong

EX 01Scheduling problem (spec over recipe)

Schedule these 9 meetings into next week (Mon–Fri, 9:00–17:00, Phnom Penh time). HARD CONSTRAINTS: - Dara: unavailable Mon; no meetings after 15:00 any day - The 2 vendor calls need US-East overlap → 20:00–22:00 local only, and those are the ONLY meetings allowed outside 9–17 - Design review must happen before sprint planning - No one gets >3 meetings/day PREFERENCES (in priority order): 1) cluster meetings to keep mornings free, 2) Friday afternoon empty. OUTPUT: a day/time table, then any constraint you could not satisfy (say so explicitly rather than silently bending one), then the one assumption most likely to be wrong.

→ Why it works

Every constraint is stated once, priorities are ordered, and impossibility has an explicit escape hatch. The model's own search handles the combinatorics better than your 200-word recipe would.

EX 02Choosing effort level per task

Route by task: 1. Email → {category, urgency} extraction: fast non-reasoning model, schema mode, few-shot examples. (Pattern matching — reasoning adds latency and cost, not accuracy.) 2. Weekly 'which 3 product bets should we kill' memo: reasoning model, high effort. Prompt = full context pack (metrics tables, constraints, definition of 'kill-worthy') + 'list the strongest argument against each recommendation'. 3. Contract-clause risk triage: reasoning model, medium effort, with a rubric of what counts as high-risk — plus 'quote the clause for every flag' as the verification hook.

→ Why it works

Effort is a dial, not a virtue. Extraction gets speed, genuine deliberation gets budget, and every reasoning task ships with its verification hook.

03 Key Points

01Skip 'think step by step' and recipe-style reasoning instructions — spec the problem, not the thought process

02Spend your prompt tokens on completeness: every fact, constraint, and edge case in one place — reasoning amplifies spec quality, good or bad

03Match effort to task: high reasoning for genuine multi-step problems; low/none for extraction, formatting, and lookup (slower + pricier for nothing)

04Define 'correct' precisely: acceptance criteria, tie-breaking rules, what to do under ambiguity

05Add verification hooks: 'list the assumptions your answer depends on' or 'give the counterexample that would falsify it'

06Don't ask it to show all its work in the answer — ask for the conclusion plus the load-bearing reasoning only

04 Model-Specific Notes

Claude's extended thinking has a budget you set per call. Peek at the thinking output during development to find spec gaps (watch it wrestle with an ambiguity you can fix), but write prompts that don't depend on reading it in production.

05 For Your Role

For a genuinely hard decision (contract, big purchase), give the reasoning model everything — numbers, constraints, what matters to you, ranked — then ask for the recommendation *plus what would change it*.

Next: Prompt Compilers & DSPy →