How to Prompt Gemini 3 Deep Think for Reliable Math Answers
A reasoning model is not a magic wand. Gemini 3 Deep Think can produce a beautifully argued wrong answer when the prompt hides an assumption or sets up the problem ambiguously. The thinking is only as good as the question.
These structures are what separate a reliable result from a confident guess.
Why do reasoning models still fail?
Two reasons. First, ambiguity: if a term in your problem has more than one reading, the model picks one silently. Second, unstated constraints: the model optimizes for what you said, not what you meant. Deep Think reduces arithmetic and logic slips, but it cannot read your mind about units, edge cases, or which variable is fixed.
Four prompt structures that improve accuracy
- Define the output. “Give the answer as a single number with units, then show the derivation.”
- List all constraints explicitly. Number them so the model can refer back.
- Ask for a self-check. “After deriving, verify by substituting the answer back in.”
- Request failure modes. “List two ways this reasoning could be wrong.”
The self-check instruction is the highest-value one. Models that verify their own answer catch a meaningful share of errors that a single pass misses, and the check costs you nothing but a few more reasoning tokens.
How do you handle units and edge cases?
State the unit system up front and forbid silent conversion. “All quantities in SI units; do not convert unless asked.” For edge cases, ask explicitly: “Check the behaviour at zero and at the boundary values.” A reasoning model left to its own devices will solve the typical case and skip the degenerate ones.
Should you give it a worked example?
For structured problems, yes. One solved example of the same shape, with the answer, cuts formatting errors and shows the model the level of detail you want. Do not give an example of the exact problem you are asking — you will get the example back. Use a different instance of the same pattern.
How do you verify a long derivation?
Break it. Ask for the derivation in numbered stages, then send a later stage back with “start from this line and continue.” If the model cannot reproduce its own middle steps, that is the weak link. Independent checks — plugging the answer back in, or solving a simplified version by hand — beat re-reading the prose.
When is all this effort not worth it?
For quick estimates or when an approximate answer is fine. If you need “roughly how many,” a fast model gives you that in a second. Reserve the careful prompt scaffolding for answers you will act on or publish.
Key Takeaways
- Reasoning models fail on ambiguity and unstated constraints, not just arithmetic.
- Define the output format, list constraints, and request a self-check.
- Fix the unit system and ask it to test boundary values explicitly.
- Give a worked example of the same shape, not the exact problem.
- Verify long derivations by breaking them into stages and checking independently.
