Prompting for Analysis & Research

Structuring prompts for rigorous, evidence-based analytical work

Analysis prompts need a different structure than writing prompts

Writing tasks reward voice and fluency. Analytical tasks reward rigor, evidence, and calibrated confidence — and if you prompt them the same way, you’ll get confident-sounding prose that’s less reliable than it reads. This lesson covers the structural techniques that push Claude toward genuinely rigorous analysis rather than analysis-shaped writing.

Separate claims from evidence, explicitly

The single biggest lever for analytical rigor is requiring every claim to be tied to its supporting evidence, in the response itself — not just as a general “be accurate” instruction, but as a structural requirement.

Analyze whether we should expand into the mid-market segment based
on this sales data. For every claim in your analysis, use this
format:

CLAIM: [the claim]
EVIDENCE: [specific data point supporting it]
CONFIDENCE: [high / medium / low, and why]

Do not include a claim you cannot support with a specific data
point from the material provided.

This format does two things: it prevents unsupported assertions from hiding inside fluent prose, and the confidence rating forces Claude to distinguish between what the data clearly shows and what it merely suggests.

Ask for the counter-case, not just the case

Left to a single pass, analysis tends to build toward whatever conclusion feels most supported by the framing of your question — a mild form of confirmation bias baked into how the prompt was asked. Explicitly requesting the strongest counter-argument counteracts this.

After your analysis, add a section: "Strongest case against this
conclusion." Steelman it — write the best version of the opposing
argument, not a weak version you can easily dismiss.

Separating fact, inference, and speculation

Especially with research synthesis, it’s easy for an inference to read exactly like a stated fact once it’s in fluent prose. Ask Claude to explicitly tag the type of each statement when precision matters.

Label each statement in your summary as:
[FACT] — directly stated in the source material
[INFERENCE] — a reasonable conclusion drawn from the source
[SPECULATION] — plausible but not directly supported

This is tedious for casual work, but for anything going into a decision memo, it’s the difference between a document you can defend and one where you can’t tell what’s solid.

Calibrating confidence instead of hedging everything (or nothing)

By default, Claude can lean toward uniform confidence — either hedging everything mildly, or stating conclusions with more certainty than the underlying evidence supports. Explicitly asking for calibrated, variable confidence per claim (as in the CLAIM/EVIDENCE/CONFIDENCE format above) produces a much more useful and honest document than a blanket instruction to “be careful.”

Structuring multi-source research synthesis

When synthesizing multiple sources with signposting (Lesson 3), add an explicit instruction on how to handle disagreement between sources — this is a common analytical blind spot if left unaddressed.

If sources disagree on a point, don't average them into a vague
middle answer. State the disagreement explicitly: which sources
say what, and if possible, why they might differ (different
methodology, different time period, different sample).