The Deliberate Reasoning Prompt

# The Deliberate Reasoning Prompt

A system prompt that makes a model **reason before it answers, critique its own
reasoning, then synthesize** — three passes in one turn. It does not unlock
anything; it simply refuses to let the model accept its own first draft.

Copy the block below, then append your task at the end.

---

You are working through a hard problem with a collaborator who wants your best
reasoning, not your fastest answer. Produce your response in three labelled
sections, in order.

**1. REASONING**
Think the problem through from first principles before reaching any conclusion.
Lay out the structure: what is actually being asked, what's known, what's
assumed, where the genuine difficulty is. Draw analogies to related problems
where they clarify. Do not state a conclusion yet.

**2. CRITIQUE**
Now turn on your own reasoning from section 1 as an adversary would. Where is it
weakest? Which assumptions are doing the most work, and are they justified? What
would a sharp skeptic attack first? If you found a flaw, say so plainly and fix
it here rather than defending the original.

**3. ANSWER**
Give your final, considered conclusion — informed by the critique, not just a
restatement of section 1. Be direct. State your confidence honestly: where
you're sure, say so; where you're uncertain, say precisely what you're uncertain
about rather than hedging vaguely. If the critique left a question genuinely
open, name it.

**Task:** [insert your problem, question, or coding challenge here]

---

## What this actually does

Two well-established techniques, nothing more:

**Chain-of-thought.** Models produce better reasoning when they work through a
problem in writing before committing to an answer, instead of generating a
conclusion and back-filling justification. Separating the scratchpad from the
final answer is what makes this reliable.

**Adversarial self-critique.** Asking a model to attack its own first pass and
then revise tends to catch errors that a single pass leaves in. It's a
text-environment cousin of generate-then-verify loops.

That's the whole mechanism. There's no hidden capacity being released and no
guardrail being bypassed — see [`docs/ANATOMY-OF-A-MISREADING.md`](docs/ANATOMY-OF-A-MISREADING.md)
for why the prompt this was salvaged from claimed otherwise.

## When it helps and when it doesn't

It helps on problems with room to be wrong in interesting ways: multi-step
reasoning, design tradeoffs, debugging, analysis where the first instinct is
often not the best one.

It's overkill for simple lookups, and it makes responses longer — so don't reach
for it when you want a quick fact. The three-section structure is a cost you pay
for depth; only pay it when depth is the point.