Thinking Modes and Reasoning Models: When to Use Them and How to Prompt Them Differently

What thinking modes and reasoning models do

A reasoning model works the problem before it answers you. A standard model produces its reply in a single pass, committing to the first word before it has considered the last. A reasoning model first generates a private working-out pass — it tries an approach, checks the intermediate result, and abandons paths that do not hold up. Only then does it write what you see.

Anthropic’s engineering documentation puts the reason plainly: a model answering in one pass has to get everything right on the first try, with no scratch work and no changing course halfway through. Fine for a thank-you note. Not fine for reconciling three versions of an expenses policy that were written by different people in different years.

Fast thinking and slow thinking

Psychologists have a useful pair of labels for this and the AI industry borrowed them wholesale. System 1 is fast, automatic and pattern-driven — the way you read a familiar word or judge a colleague’s mood across a room. System 2 is slow and effortful, the mode you drop into for long division, or for a contract clause you cannot afford to skim.

Standard models are close to pure System 1. They are extraordinarily good at producing the plausible next thing, which is exactly why they can be confidently wrong on a problem that needed working out. Reasoning models add a System 2 pass on top, and the trade never changes: more time and more cost, against fewer confident mistakes on problems with many moving parts.

Where the working-out goes

You are billed for that hidden pass whether or not you ever read it, and what you can see is a summary rather than the raw internal reasoning. No major tool exposes the unedited version, and Anthropic’s documentation on how thinking works says so directly.

In practice that means one thing. You cannot audit the reasoning the way you would audit a colleague’s spreadsheet — you can read a tidied description of the route it took. Useful for spotting a wrong turn. Not proof the destination is correct.

The toggle is quietly disappearing

Across every major tool, thinking modes and reasoning models have moved from something you switch on to something you occasionally switch off. Automatic routing now runs by default, and the free tiers have caught up faster than most people realise.

Two details in that table are worth more than the rest. On ChatGPT, an automatic switch into Thinking does not draw down your weekly Thinking allowance — only choosing it manually does. And on Claude’s current models the deliberation is already running; the model decides how deep to go rather than waiting for you to ask.

Google splits the difference. Standard and Extended thinking levels went out to everyone across web and mobile, while Deep Think, which Google DeepMind describes as its most specialised reasoning mode, stays behind the Ultra subscription.

Automatic routing is right most of the time, and wrong in a predictable direction: it under-thinks. A request that looks casual gets handled as casual, even when the work sitting behind it is not. Three situations reliably call for overriding upward:

• The material is long and the contradictions are buried rather than obvious
• You want a judgement across several sources, not a summary of one
• The answer goes to someone senior without you re-checking it first

Why your old prompt now backfires

The scaffolding that used to rescue a weak model now gets in a strong one’s way. Older systems needed to be walked through their own reasoning out loud, which is where “let’s think step by step” came from — it forced a model that would otherwise leap to an answer to show intermediate work.

Reasoning models already do that internally. When you also hand them a rigid sequence of steps, you are not adding rigour; you are overwriting a process that was better than the one you supplied. The model follows your route instead of finding its own, and your route was written by someone who had not yet seen the documents.

The second version is shorter and it is doing more work. It names the destination, hands over the raw material, and defines what a finished answer looks like. Everything about how to get there is left to the model, which is the part it is genuinely better at than you are.

Three habits are worth deleting by name, because each one was sound advice once:

• The invented persona. “You are a world-class consultant with 20 years of experience” adds nothing the model does not already have, and narrows the answer into something performatively expert.
• The reasoning instruction. Any variant of “think step by step” now competes with a better internal process. You are asking for working it was already doing.
• The effort plea. “Be thorough”, “be careful”, “do not miss anything” — these change tone rather than diligence. They read as pressure and tend to produce padding.

This is the same argument we make about structured thinking beating prompt tricks, arriving from a new direction. It also explains why old advice ages badly — a technique tied to a model’s weakness stops working the moment that weakness is fixed. If you want the longer history, our guide to what prompt engineering actually means covers where these habits came from.

The Destination Brief, step by step

The Destination Brief is the replacement habit we teach at PromptPeakAI: state where the output needs to land, hand over everything relevant, mark the hard lines, and say what finished looks like. Four parts, no step-by-step choreography.

Write the outcome you need in one sentence, as if briefing a capable colleague who will not be able to ask you a follow-up question. “Find the contradictions” is a destination. “Read each document, then list the rules, then compare them” is a route, and it is the part you should stop writing.

Give it the actual documents rather than your summary of them. Summarising first throws away exactly the detail the deliberation pass is there to catch — a clause you skimmed past is often the clause that matters. This is where the honest data question arrives, so settle it before you paste.

Constraints are limits that would make an answer wrong or unusable — a jurisdiction, a budget ceiling, a policy that cannot be contradicted, a length the format demands. They are not stylistic preferences dressed up as rules, and they are not instructions about how to reason.

Describe the finished artefact: the shape, the columns, the length, what must appear in it. This is the single highest-value line in most requests, and the one people most often leave out.

Goal: [the outcome you need, in one sentence]

Inputs:
[paste the full documents, data or notes here — do not summarise them first]

Constraints:
- [a genuine hard limit, e.g. must comply with our 2026 travel policy]
- [another, e.g. figures in GBP only]

Done when:
- [what the finished output contains, e.g. a table with these four columns]
- [any coverage requirement, e.g. every contradiction found, not just the clearest]
- Flag anything you are uncertain about rather than smoothing it over.

Here is the same skeleton filled in for the job that shows the difference most clearly — a reconciliation across documents that disagree with each other.

Goal: Identify every point where these three documents contradict each other on expense approval, and tell me which version currently governs.

Inputs:
[paste the operations manual extract]
[paste the expense policy]
[paste the remote work guidelines]

Constraints:
- Only flag genuine contradictions, not differences in wording.
- Where a document is undated, say so rather than guessing its order.

Done when:
- A table with: topic, what each document says, the nature of the conflict, which version appears authoritative.
- A short list of the contradictions that carry financial or legal risk.
- An explicit note of anything you could not resolve from the text provided.

If your work involves the same brief every week, it is worth saving the skeleton somewhere you can reach it — our free AI work templates include a version you can adapt. The same structural logic sits behind the Logic System we use across the site, and behind chaining several requests together when one brief genuinely cannot carry the whole job.

When the slow mode is not worth it

Deliberation costs you two things: time, and the tighter allowance that manual deep-mode picks draw from. Spending both on a task that never needed them is the most common waste in day-to-day AI use, and it is entirely avoidable.

The pattern that works for most professionals is a split shift. Do the hard thinking once in the deep mode — the analysis, the reconciliation, the plan — then drop back to the fast default for every round of editing that follows. Refinement is rarely the part that needs deliberation.

Here is the analysis you produced earlier: [paste it]

Change only the following and leave everything else exactly as written:
- [change one, e.g. cut the third section to a single paragraph]
- [change two, e.g. replace "stakeholders" with "the finance team" throughout]

Do not re-analyse, re-order or improve anything I have not asked you to change.

Which work actually justifies it

A short list of task types accounts for most of the genuine wins. They share a shape: several sources, an answer that is not visible from any one of them, and a real cost to getting it wrong.

• Reconciling documents that disagree. Policy versions, contract revisions, a handbook against the practice that quietly evolved around it. Operations and HR leads hit this constantly.
• Stress-testing a plan or a budget. Several variables moving at once, where the useful answer is which combination breaks first.
• Root-cause work on a messy incident. Scattered logs, contradictory accounts, a timeline nobody wrote down — separating causes from symptoms.
• Structured evaluation against fixed criteria. Vendor proposals or applications scored on a matrix, where consistency across items matters more than eloquence on any one.

Most procurement and finance professionals will tell you the same thing about that last one: the value is not speed, it is that the twentieth submission gets assessed as carefully as the first. Consistency is where human attention fails earliest, and it is the thing a deliberation pass genuinely protects.

Research-style work is the honourable exception, because the deliberation is the product rather than an overhead. If that is your task, the purpose-built modes will serve you better than a long chat: ChatGPT’s Deep Research mode, Gemini Deep Research, and Copilot’s Researcher agent each exist for exactly this.

When reasoning still gets it wrong

A reasoning model that has thought carefully still produces confident, well-organised, plausible text when it is wrong. That is the failure mode worth internalising, because the deliberation makes the output look more trustworthy while doing nothing for the thing you actually needed checked.

Most of the remaining problems are self-inflicted and quick to fix once you recognise them.

The third row is the one that catches people out, and it is worth a deliberate habit rather than good intentions. Our guide to why AI invents things that sound right covers the mechanism; the practical response is a separate verification pass, run cold.

Below is an analysis you produced. Do not improve it or rewrite it.

[paste the analysis]

For every factual claim in it — names, figures, dates, quoted policy wording — tell me:
- the exact sentence in my source documents that supports it, or
- that no sentence in the source supports it.

List anything unsupported first. Do not fill gaps from general knowledge.

Thinking modes and reasoning models have quietly stopped being a setting you manage and become a default you occasionally override. That is a smaller skill than the older guides describe, and a more useful one: knowing when your judgement should beat the router’s, and writing a brief good enough to deserve the extra minute.

The rest follows from a single change of instinct. Give the model the whole problem, the real material, and a clear picture of what finished looks like — then get out of the way. Wanting to over-explain is the hardest habit to drop, and dropping it is most of the improvement.