How to Write Prompts That Work Best in Gemini

A operations coordinator at a regional healthcare clinic once asked Gemini to "review this patient intake form for issues" and got back a single line saying the form looked fine. It didn't. It was missing an entire required consent section that the compliance team caught two weeks later during an audit. The form wasn't the problem — the prompt was. "Review this for issues" gave Gemini no criteria to check against, so it did the AI equivalent of a glance instead of an audit.

That failure is a good entry point into how to write prompts for gemini, because it's not really about Gemini being unreliable — it's about how much weight a vague prompt puts on the model to guess what "issues" means, and how often that guess lands on "nothing's wrong" by default. This guide walks through the specific structural habits that fix this, with copy-paste templates you can adapt today.

Quick-Start (Copy This Right Now)

[Task] with these explicit criteria to check against: 
[criteria 1], [criteria 2], [criteria 3]

For each criterion, tell me explicitly whether it's met, not met, 
or unclear from what I've given you. Don't just tell me if 
something looks generally fine.

What this does: Naming explicit criteria and forcing an explicit pass/fail/unclear verdict on each one is the single highest-leverage change you can make to a review-style prompt. It's the difference between "this looks fine" and an actual audit.

This structure works because it removes the ambiguity about what "good" means from the model's side of the conversation and puts it back on your side, where it belongs. You're the one who knows what the intake form actually needs to contain. Gemini's job is to check your document against your criteria accurately, not to independently guess what criteria might matter — and when a prompt implicitly asks it to guess, it will, and that guess defaults toward the reassuring answer more often than not.

Understanding the Variables

Gemini, like most large language models, defaults to being agreeable and confirmatory when a prompt doesn't push it to look for problems specifically. Ask an open-ended "does this look okay" question and you'll often get an open-ended "looks okay" answer, because that's the path of least resistance when no specific failure criteria are defined.

The fix isn't asking Gemini to "be more careful" — vague instructions to be careful don't change what the model is checking, only its tone while doing it. The fix is giving it a specific list to check against, the same way you'd hand a human reviewer a checklist instead of just saying "take a look."

This pattern shows up constantly outside of formal document review too. Ask "is this email okay to send" and you'll usually get reassurance. Ask "check this email against these criteria: professional tone, no typos, includes the meeting time and location, doesn't overpromise on the deliverable date" and you get an actual evaluation instead of a rubber stamp. The structure of the request, not the subject matter, is what determines whether you get a real check or a polite non-answer.

⚡ Pro tip: For any review or audit-style task, explicitly ask for "unclear" as a third option alongside pass/fail. Without that option, Gemini sometimes forces itself into a binary "fine" or "not fine" verdict on things that genuinely can't be determined from the information you've provided, which can look like false confidence.

Step-by-Step: Structuring a High-Stakes Prompt

Step 1 — Define what "success" looks like before you ask for anything. For the intake form example, success meant a specific list: patient identifying information present, consent language present, insurance fields complete, and a signature line. Vague success criteria produce vague reviews.

I need this [document/form/piece of content] to meet these specific 
requirements: [list them explicitly, even if they seem obvious to 
you].

Step 2 — Ask for the check, not just the fix. Requesting a review and requesting a rewrite are different tasks, and combining them tends to produce a confident-sounding rewrite that quietly glosses over a problem instead of flagging it clearly.

First, check this against the requirements above and tell me 
specifically what's missing or wrong. Don't fix anything yet — 
just the audit.

⚠️ Common mistake: Asking for a review and a fix in the same request. When you ask Gemini to "check this and fix any issues" in one breath, it tends to jump straight to fixing what it notices and skip explicitly telling you what was wrong in the first place — which means you never actually learn what the gap was, and you can't verify the fix addressed the real problem.

Step 3 — Verify anything high-stakes with a second, differently-worded pass. For anything genuinely consequential — compliance documents, financial figures, medical or legal content — run the check twice with slightly different phrasing, and treat disagreement between the two passes as a signal to look closer yourself rather than trusting either answer blindly.

Pro-Level Variations

For data-heavy tasks, add a source-grounding instruction: "Only use information from the document I've provided — if you need to make an assumption not stated in the document, flag it explicitly rather than filling the gap silently." This matters enormously for anything where an invented detail could be mistaken for a real one.

For long-document analysis, break the criteria into stages rather than asking for everything at once: "First, list every section in this document. Then I'll tell you which sections to check against which criteria." This works better than a single mega-prompt because it forces Gemini to actually map the document's structure before evaluating it, rather than skimming for relevant-looking chunks.

For anything involving numbers or calculations, add an explicit instruction to show its work: "Show the calculation for each figure, not just the final number." I'm not 100% sure why, but forcing the intermediate steps to be visible seems to reduce arithmetic slip-ups noticeably compared to asking for a final answer alone — possibly because it's harder for an error to hide when the reasoning that produced it is right there to check.

For recurring reviews of the same document type — the clinic's intake forms, for example, get revised periodically — save the criteria list itself as a reusable template rather than retyping it each time. A compliance coordinator at a different healthcare practice keeps a standing prompt with the full checklist already written out, so reviewing a new form revision takes one paste instead of reconstructing the criteria list from memory each time, which is exactly the kind of place small omissions creep back in.

A Word on Trusting the Output

None of this makes Gemini infallible, and it's worth saying plainly: even a well-structured prompt with explicit criteria can miss something a domain expert would catch, particularly for genuinely high-stakes content like medical, legal, or financial documents. Treat a well-structured review prompt as a strong first pass that catches the obvious gaps and frees up a human reviewer's attention for the subtler judgment calls, not as a replacement for that human review entirely. The clinic's fix wasn't "trust Gemini more" — it was "give Gemini a real checklist, and still have a compliance person do a final read before anything goes out."

Troubleshooting Common Issues

If Gemini keeps giving you a generically positive review no matter how you phrase the request, check whether you've actually specified failure criteria at all, or just asked it to "check" something in the abstract. Abstract review requests get abstract, usually favorable, answers.

If it's flagging things that aren't actually problems, you may have given criteria that are too rigid for legitimate variation — loosen the criteria slightly and ask it to note anything ambiguous rather than treating every deviation as a hard failure.

If the same prompt produces different verdicts on different attempts, that's your signal to not rely on a single pass for anything consequential — this is exactly why the second-pass verification step matters most for high-stakes content, not a sign that something's uniquely wrong with your prompt.

Your Turn

The clinic's intake form mistake wasn't really about Gemini's capability — a version of the same prompt with explicit criteria almost certainly would have caught the missing consent section. The gap was entirely in how the request was structured.

This is a pattern worth internalizing beyond just document review: nearly every time an AI tool's output feels disappointingly generic or unreliable, the more productive question isn't "is this tool good enough" but "did my prompt actually specify what a good answer would look like." It's a less satisfying diagnosis than blaming the model, but it's the one that's actually fixable on your end, today, without waiting for the next model release.

Once you've built a review-style prompt with the right criteria for your specific documents or content, that's exactly the kind of thing worth saving rather than rebuilding from memory every time — I keep mine in PromptABCD with the criteria list already filled in, so pulling up "intake form review" starts from a tested checklist instead of a blank, vague request.