Fluent prose reads as confident prose. When Claude summarizes a contract or answers a policy question, the sentence built on solid evidence and the sentence built on a plausible guess arrive in the same polished register — unless your prompt asks the model to distinguish them. Calibrated uncertainty (an answer whose expressed confidence roughly matches how well-supported it is) is a promptable behavior, and for enterprise outputs that feed decisions, it belongs in the system prompt. These techniques are plain Messages API prompting and work the same on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and Claude Platform on AWS.
Ask for uncertainty explicitly — and explain why it's welcome
By default, assistants are tuned toward helpfulness, and an unqualified answer often reads as more helpful than a hedged one. Anthropic's prompting guidance says to explicitly request the behavior you want rather than implying it, and to explain why the instruction matters so Claude can generalize. Applied here:
When you answer:
- If the provided documents directly support a claim, state it
plainly and name the source.
- If you are inferring beyond the documents, mark it: "Inference:".
- If you cannot verify something, say "I can't verify this" rather
than guessing.
An accurate "I'm not sure" is more valuable to us than a fluent
guess — analysts act on these answers, and a wrong-but-confident
answer costs far more than a flagged gap.That final explanation is not decoration. It tells Claude which way to lean on every borderline case you didn't enumerate, and it counteracts the pull toward unqualified fluency.
Tie confidence to evidence, not to tone
The most robust hedging pattern grounds every claim in a source the model can point to. Phrases like "based on the provided text" are useful precisely because they scope the claim: they say this is what your documents state, not this is true of the world. If your inputs are documents, you can go further than prompting — the Messages API citations feature ("citations": {"enabled": true} on document blocks) returns cited_text spans tied to specific documents, turning "based on the provided text" from a stylistic hedge into a checkable reference. Anthropic's docs note this is also efficient: cited text does not count toward output tokens, unlike prompting Claude to quote passages.
Keep hedging from swallowing the answer
Over-corrected prompts produce a different failure: every sentence wrapped in qualifiers, disclaimers stacked three deep, and a reader who can't find the answer. Set both edges of the behavior. Instruct Claude to answer plainly when evidence is strong, hedge only where support is genuinely thin, and never use hedging language as filler. A few well-chosen examples help enormously here — Anthropic recommends 3–5 diverse examples in <example> tags, and for calibration the diversity should span the confidence spectrum: one firmly supported answer stated flat, one partially supported answer with a scoped hedge, one unanswerable question refused cleanly. The model learns the gradient, not just the extremes.
Make confidence machine-readable when it gates decisions
If downstream logic routes on confidence — auto-approve high-confidence extractions, queue low-confidence ones for human review — don't parse hedges out of prose. Ask for a discrete field: "confidence": "high" | "medium" | "low" plus a short "basis" string explaining the rating. You can request this via prompt-only JSON patterns or enforce the shape with structured outputs (output_config with a JSON schema), which is generally available on the Claude API, Claude Platform on AWS, Bedrock, and Vertex AI, and in beta on Foundry. One caveat from the official docs: citations cannot be combined with structured outputs in the same request, so evidence-linked answers and schema-enforced confidence fields need to be separate calls.
Two honest limits to note. First, self-reported confidence is the model's judgment, not a statistical guarantee — treat the labels as a triage signal and validate them against outcomes on your own data. Second, thresholds are yours to set: run a sample through the pipeline, check how often "high" is actually right for your task, and tune the routing accordingly. Anthropic's prompt-engineering overview treats empirical testing against explicit success criteria as the prerequisite for all of this.
Where to go next
Calibrated uncertainty pairs naturally with reducing hallucination through prompt design and grounding prompts for RAG applications. For the regulated-industry version of this discipline, see prompting in sensitive domains.