ChatGPT vs Claude vs Gemini is a natural question when you are deciding where to paste a prompt. PromptByJob publishes role based AI prompts for accountants, software engineers, sales reps, marketers, data analysts, nurses, teachers, HR managers, lawyers, content writers, customer support agents, and real estate agents, and each prompt includes specific rules about structure, length, and tone. I wanted to see how closely three popular models follow those rules when given the same prompt.
I had opinions, and so does everyone in this space. But I did not want to publish another vague comparison post that says “Claude is great at writing and ChatGPT is great at everything” without showing the work. So I picked twelve prompts from my site, one from each profession category, and ran each one through ChatGPT, Claude, and Gemini with the same wording, the same formatting instructions, and the same output requirements. I used the first response only, with no retries, and saved each one so I could compare them.
This article walks through what I found, why instruction following matters for prompts with strict requirements, and which model did what I asked instead of what it assumed I wanted.
Why I Tested Instruction Following Instead of Just Answer Quality
Answer quality is the usual way to compare models, but it is not the only one. For job specific prompts with strict rules, a useful question is whether the model followed them. If a nurse asks for a patient handoff note in a specific format, or an accountant asks for a month end variance commentary with three required sections, what matters for that request is whether the response respected the structure, the word count, the tone, and the constraints written into the prompt.
This is usually called instruction following. It is related to, but not the same as, how well a model reasons. A model can write a strong answer and still miss a formatting rule such as a word limit or a section order. That gap is what I wanted to look at, because it affects how much checking and editing an output needs before you can use it.
What People Are Actually Searching For
Job specific prompts usually come with rules attached: a required structure, a length limit, a banned phrase, or a tone. A response that ignores even one of those rules can mean extra editing, or a deliverable that does not match a template. For that reason, this test checks rule by rule whether each model did what the prompt asked, instead of judging which answer sounds best.
How I Set Up the Test
I used twelve prompts from my own site, one per profession. Each prompt contained at least three explicit instructions, such as a word count range, a section order, a tone requirement, or a formatting rule like bullet points only. I ran each prompt once in ChatGPT (GPT 5.5), once in Claude (Sonnet 5), and once in Gemini (Gemini 3.1 Pro), with custom instructions and memory turned off, and I recorded the first response from each.
Limits of This Test
Each prompt was run once per model, so these results show one response per prompt and not an average. Models can respond differently to the same prompt on different runs, and I scored the responses myself, so another person might score some of them differently. The results describe these twelve prompts and the model versions named above. They do not rank the models in general.
The Twelve Categories I Tested
I ran prompts across software engineering code review, accounting month end close, sales cold outreach, marketing campaign briefs, data analyst report summaries, customer support response templates, content writing outlines, real estate listing descriptions, HR policy drafts, legal case brief summaries, nursing patient handoff notes, and teacher lesson plan outlines.
ChatGPT vs Claude vs Gemini: Round by Round Results
Software Engineering: Code Review Prompt
Review the following code for security issues only. Structure your response in exactly three sections: a short summary, a severity ranked list of issues from highest to lowest risk, and a suggested fix block for each issue. Do not explain any single issue in more than two sentences. Do not include style or performance feedback, security only.
The prompt asked for a security focused code review with three fixed sections, a summary, a severity ranked list of issues, and a suggested fix block, plus a hard rule of no explanations longer than two sentences per issue.
Claude followed the section order exactly and kept every explanation inside the two sentence limit. It also ranked the issues by severity instead of listing them in the order they appeared in the code. ChatGPT gave a strong technical answer but ignored the two sentence cap on two of the five issues, expanding them into paragraph length explanations. Gemini stuck to the format but gave thinner security reasoning and flagged fewer issues than the other two.
Accounting: Month End Variance Commentary
Write a month end variance commentary in first person business language. Structure it in exactly three paragraphs, one covering revenue variance, one covering expense variance, and one covering the overall margin impact. Never use generic phrases like due to various factors or a number of reasons. Every variance must be tied to a specific, named driver.
This prompt required first person business language, a strict three paragraph structure, and a ban on generic phrases like “due to various factors.” ChatGPT did the best job here of the three, keeping tight paragraphs and specific language. Claude also followed structure well but occasionally drifted into a slightly more formal tone than requested. Gemini broke the paragraph count, delivering four paragraphs instead of three, and used one of the exact banned phrases the prompt told it to avoid.
Sales: Cold Outreach Sequence
Write a three email cold outreach sequence for the offer described below. Each email must stay under 600 characters including spaces. After the three emails, add a short reasoning block explaining why each email is structured the way it is and why it sits in that position in the sequence. Do not restate the email content in the reasoning block, explain the strategy only.
The prompt asked for a three email sequence with a character limit per email and a short reasoning block explaining why each email was structured that way. Claude handled the reasoning block best, explaining its structural choices instead of restating the email content. ChatGPT stayed almost exactly within the character limits but gave a thinner reasoning explanation. Gemini exceeded the character limit on the second email by a wide margin, which breaks the main length rule in the prompt.
Marketing: Campaign Brief
This one asked for a structured brief with five labeled sections and a specific instruction to avoid marketing buzzwords like “synergy” or “leverage.” All three models respected the section labels. ChatGPT slipped in one buzzword despite the explicit ban. Claude avoided every flagged word and kept the brief the most concise of the three. Gemini was competent but generic, reading like a template rather than a brief built around the specific product details I gave it.
Data Analysis: Report Summary
The test here was whether the model would stick to describing only the data provided instead of inventing additional context or assumptions. This is where model behavior diverged the most. ChatGPT added assumptions that were not in the data, filling gaps with plausible sounding guesses. Claude stayed the most disciplined, explicitly flagging when something could not be determined from the given numbers instead of guessing. Gemini was in between, mostly accurate but with one instance of adding a trend claim the data did not support.
Customer Support: Response Template
Write a customer support response for the issue described below. The tone must be warm and helpful but not apologetic. Do not use the words sorry, apologize, or unfortunately anywhere in the response. Acknowledge the issue in one sentence, then move directly into the solution.
The prompt simulated a support style guide that bans apology language. Gemini and ChatGPT both slipped into apologetic wording despite the instruction. Claude was the only one that avoided apology language in its response.
Content Writing: Article Outline
This test measured whether the model would respect a strict heading count and depth requirement, specifically five H2 headings with two H3 subheadings each. ChatGPT and Claude both hit the structure precisely. Gemini added an extra H2 that was not requested, which is a small deviation but still a deviation from the instruction.
Real Estate: Listing Description
The instruction here banned superlatives like “stunning” or “amazing” and required the description to stay under a strict character count for MLS compatibility. All three models respected the character count. Claude and ChatGPT avoided the banned superlatives completely. Gemini used one banned word.
HR: Policy Draft
This prompt required neutral, legally cautious language and a specific instruction not to use absolute terms like “always” or “never.” Claude was the most careful here, consistently using conditional language. ChatGPT slipped once with an absolute phrase. Gemini did reasonably well but read more like a general policy template than something tailored to the scenario I described.
Legal: Case Brief Summary
The test measured whether the model would stick to the facts I provided instead of drawing legal conclusions I did not ask for. ChatGPT stayed within the facts given. Claude did the same. Gemini added an interpretive conclusion that went beyond the scope of the prompt. This test checked whether each model stayed within its instructions, not whether the legal content is accurate, and it is not legal advice.
Nursing: Patient Handoff Note
Write a patient handoff note using strict SBAR format, situation, background, assessment, and recommendation as four separate, clearly labeled sections. Do not merge assessment and recommendation into one section. Do not add any commentary outside the four sections. Keep each section to two or three sentences.
This prompt required a rigid SBAR structure with four labeled sections: situation, background, assessment, and recommendation, with no extra commentary. Claude and ChatGPT both followed the structure precisely. Gemini merged the assessment and recommendation sections instead of keeping them separate, which breaks the four section format the prompt asked for.
This comparison checked formatting only. It did not evaluate whether the clinical content is accurate or appropriate, and nothing in this article is medical advice. If you use AI to draft handoff notes, follow your employer’s policies on patient information, do not enter identifiable patient details into a tool your employer has not approved, and have the text checked by a qualified professional before relying on it.
Teaching: Lesson Plan Outline
Create a lesson plan outline for the topic described below. Break it into labeled time blocks that add up to exactly fifty minutes total, no more and no less. Show the minutes allotted next to each block. Include a warm up, main instruction, guided practice, and a closing check for understanding.
The final test asked for a lesson plan with a fixed time allocation per section, adding up to exactly fifty minutes. ChatGPT and Claude both did the math correctly and respected the time blocks. Gemini’s time allocations added up to fifty three minutes, a small arithmetic slip that would actually matter to a teacher planning a real class period.
The Final Scoreboard
Counting the deviations described in the rounds above, Claude followed every explicit instruction on eleven of the twelve prompts. Its one slip was a slightly more formal tone than requested on the accounting prompt. ChatGPT followed every instruction on seven, with slips on five prompts: the two sentence cap in the code review, a banned buzzword in the marketing brief, added assumptions in the data summary, apology language in the support response, and an absolute phrase in the HR policy. Gemini followed every instruction on three, with deviations on the other nine: a wrong paragraph count and a banned phrase in the accounting commentary, an exceeded character limit in the sales emails, an unsupported trend claim in the data summary, apology language in the support response, an extra heading in the content outline, a banned word in the listing description, an added legal conclusion, merged SBAR sections, and lesson plan time blocks that added up to fifty three minutes.
None of the three models missed every rule, and many of the deviations were small. But a response can be a usable draft and still break a rule, and if you use a detailed prompt so that you do not have to check every constraint afterward, a missed rule defeats the purpose.
Why This Pattern Makes Sense
Across these twelve prompts, ChatGPT’s slips were usually small misses, such as a banned word, an apology phrase, or a sentence cap. Gemini had the most deviations, including extra or merged sections, a missed length limit, time blocks that did not add up, and a few banned words. Claude had one documented slip, a slightly formal tone on the accounting prompt. These are observations from one response per prompt, and they may not hold for other prompts or other model versions.
None of this means one model is universally better. In this test the models differed in how literally they followed explicit instructions, and that matters more for some tasks than others. A marketer brainstorming campaign angles may not mind if a model adds an extra idea. For tasks with strict format or wording rules, such as a handoff note template or a case summary, a missed rule means more checking, and AI generated text in those fields still needs review by a qualified person.
What This Means If You Use Job Specific Prompts
In this test of twelve prompts with hard constraints, required formats, banned language, or strict structure, Claude followed the most instructions correctly, ChatGPT followed fewer, and Gemini followed the fewest. Each model was run once per prompt, so treat this as one set of observations and not a general ranking. If your prompt has strict formatting rules, check every output against the original instructions, whichever model you use.
One takeaway from this test is that a prompt with specific, checkable rules makes it easy to see where a model drifts. Whichever model you use, write constraints you can verify, such as a word limit, a section order, or a list of banned phrases, and check the output against them before you use it.
Frequently Asked Questions
Which AI model follows instructions most accurately for professional prompts?
In this test of twelve job specific prompts, Claude followed the most explicit instructions correctly. The test used one run per prompt, so it is a limited comparison and does not show how the models will perform on every prompt.
Does a smarter AI model automatically follow prompts better?
Not necessarily. Following instructions and reasoning are related but different abilities. A model can write a strong answer and still miss a specific formatting rule. This test measured instruction following only, not reasoning.
Should I use a different AI model for different jobs?
It can help to try more than one. This test did not measure which model is best for each job. It only measured how closely each followed the instructions in twelve prompts. If your prompt has strict rules, run it, check the output against each rule, and switch models if one keeps missing the same rule.
Can I trust AI output without checking it against the original prompt?
No. Every model in this test missed at least one instruction on some prompts, including Claude, which had a tone slip on the accounting prompt. Always compare the output against your original instructions before using it in real work.
Rehan Malik is a blogger and the creator of PromptByJob. He publishes articles and prompt examples about using AI tools for different tasks. He uses AI-assisted drafting in his writing process. He is not a healthcare professional, and his nursing-related content should not be treated as medical advice or evidence of clinical testing.
