Home / Techniques that change the result
Asking several times, and contradicting yourself
Asking for several independent answers and keeping the majority filters out the noise of a single answer, and asking Claude to play devil's advocate against its own answer finds errors a single pass misses, provided you always run a final synthesis to sort the valid objections from the noise.
A single answer to a hard question can be right by chance or wrong despite reasoning that seemed solid. Two techniques offset this risk without changing the prompt itself: asking several times and keeping what recurs, or having an answer critiqued by its own author.
Several independent answers, and the majority that remains
This technique, known in large language model research under the name self consistency, does not appear by that name in Claude's official documentation as of today, it remains transferable to any language model. The principle fits in one sentence: ask the same question several times, in separate exchanges, then keep the answer that comes up most often.
Question posee trois fois, dans trois echanges separes : "23 fois 17, en detail"
Réponse 1 : 391
Réponse 2 : 391
Réponse 3 : 381Two answers out of three converge on 391, which is the exact result of this calculation. The majority settles in favour of the correct result, and the isolated outlier falls away, without having to replay the full reasoning of each exchange one by one.
Devil's advocate against your own answer
Claude's documentation recommends, as part of structured research, regularly critiquing your own approach and plan (platform.claude.com, section on research and information gathering, accessed 2 September 2026). The same reflex generalises to any answer with real stakes: ask Claude to play devil's advocate against what it has just written, specifically looking for what could be wrong or incomplete in its own answer.
The final synthesis sorts, it does not vote
Neither the majority vote nor the adversarial critique replaces a synthesis step: harsh criticism can push back a correct answer just as easily as it corrects a wrong one, if nobody then sorts the valid objections from the noise. This synthesis is built like a chain of short prompts, one link that produces the answer, one link that critiques it, a final link that decides what remains. Each independent answer and each critique happen in a separate exchange, so that a previous answer does not influence the one that follows before the synthesis brings them together.
A single answer accepted against several answers confronted
A question asked once, an answer received and used with no further checking, whatever the real stakes of the topic.
The same question asked three times in separate exchanges, the majority retained, then an adversarial critique of that answer and a synthesis that sorts the valid objections.
A person asks Claude the same calculation question, in three separate exchanges, and gets three answers. Two answers give 391, the third gives 381. She keeps the answer that comes up twice.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: Across these three exchanges, the answer 391 is the one that the majority of independent draws produced, which makes it more credible than the isolated answer 381.
What this does not establish: This does not establish that 391 is guaranteed to be the mathematically exact answer, since a majority of independent samples can converge on the same error if the model makes the same kind of mistake repeatedly.
The three most common miscalibrations
- Too broad This majority definitively proves that 391 is the correct answer to this multiplication.
- Too narrow This result proves nothing at all, since asking the same question three times is not a real check.
- Beside the point This result shows that Claude answers differently depending on how a question is worded.
- Asking the same question several times, in separate exchanges, and keeping the answer that comes up most often is a general technique, not documented by that name by Anthropic for Claude as of today.
- Asking Claude to critique its own answer, specifically looking for what could be wrong, finds errors a single pass lets slip through.
- Harsh criticism can push back a correct answer just as easily as it corrects a wrong one, if no synthesis then sorts the objections retained.
- Each independent answer and each adversarial critique happen in a separate exchange, so that none influences the one before it.
Ask Claude, in three separate exchanges, a question whose exact answer you already know, note the three answers obtained, and check whether the majority converges on the correct answer.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- If you are looking for an official source on majority voting across several answers, check directly on platform.claude.com: as of today, no page documents this technique under this name for Claude, it remains a general transferable practice.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Prompting best practices, section Research and information gathering consultée le 2026-09-02