Day 6 · Get genuinely good at it

Choosing between ChatGPT, Claude and Gemini

Test the three main assistants fairly, choose one for your work and stop comparing them.

14 min readPlus one task, about 10 min

The one idea

ChatGPT, Claude and Gemini are differently shaped, not neatly ranked. They overlap on ordinary writing, and all three can be fluent and wrong. The useful difference is the surrounding fit: the documents they handle well, the applications they connect to, the controls you prefer and the limits you meet on a normal week.

ChatGPT is the broad all-rounder. Claude is particularly good at long, careful documents and writing that needs restraint. Gemini is strongest when your work already lives in Google Search, Docs, Drive, Gmail or Sheets. Those are tendencies, not laws. A model update can move a benchmark, and a free-tier limit can matter more to you than a small quality difference.

Pick one by running the same real job through the free tiers and recording what happened. Give each tool the same input, the same instructions and the same scoring rules. Then choose a default. You can use another tool later for a specific job, but you do not need three open tabs to write a weekly email.

Why it matters

People pay for two subscriptions and use one, or spend a week reading comparisons that never touch their work. The result is not better judgement. It is decision fatigue. A beginner needs a default tool that is easy to reach and good enough for the jobs that actually recur.

The differences are real, but they are not a universal podium. A long policy document may favour Claude because it holds detail and writes cautiously. A person who spends the day in Google Docs may save more time with Gemini because the document is already in the right place. Someone who wants writing, files, voice, images and a wide range of everyday features in one app may prefer ChatGPT. The best choice is the one that reduces work in your context without asking you to ignore its weaknesses.

Today is also a lesson in test design. If you give each assistant a different prompt, use different source material or score only the answer you happen to like, you are measuring your mood. Use one task, one input and a short scorecard. “Would I reuse this?” is a more useful question than “Which model won a benchmark?”

AI-generated illustration

Three tools can be different shapes without one being the universal winner. Test the fit against your work.

The tools you will use

The tool

ChatGPT

A general-purpose assistant from OpenAI, available on the web and in mobile apps.

What matters about it here

Strong everyday drafting, rewriting, explanation and structured output
File uploads, voice on supported apps and web search, with availability depending on plan and region
A broad set of built-in capabilities, which can reduce the number of separate tools you need

How it works: It is a language model wrapped in a chat product. Your message, conversation context and any supplied file or search result are placed into the model's context, and it generates a response one token at a time. File handling and web search are extra systems that feed material into the model; they do not make its unsupplied claims automatically true.
What it costs: A free tier is available with usage and feature limits. ChatGPT Plus is commonly US$20 per month, with higher limits and access to more capable features; prices, taxes and availability vary by country.
Why this one: Choose it as a default if you want the broadest general-purpose set of features in one familiar interface and your work is a mix of writing, questions, files and occasional creative tasks.
If you cannot get it: Claude or Gemini can handle ordinary writing and questions. Choose based on the work you actually do rather than assuming the name decides quality.

The tool

Claude

A general-purpose assistant from Anthropic, known for careful writing and handling long documents.

What matters about it here

Long-context document work, useful for reports, policies, transcripts and contracts
A restrained writing style that often makes uncertainty and caveats more visible
Projects and Artifacts on supported plans, for keeping reference material or building an editable output beside the chat

How it works: It uses the same broad next-token generation mechanism as other chat assistants. When you add documents, their text is processed and placed in the context the model can use. Its training and product behaviour encourage a more cautious style, but cautious wording is not proof of accuracy and it can still invent or misunderstand.
What it costs: A free plan is available with usage limits. Claude Pro is commonly US$20 per month, with more usage and access to additional capabilities; plan names and regional prices can change.
Why this one: Choose it if your recurring work is document-heavy and you care about preserving detail, a calm tone and explicit caveats more than having every extra feature in the same app.
If you cannot get it: ChatGPT and Gemini both accept documents. For a short document, the difference is usually smaller than the difference between a clear prompt and a vague one.

The tool

Gemini

Google's assistant, available as a standalone chat and through parts of Google's wider software ecosystem.

What matters about it here

Useful connections to Google Search and, where enabled, Workspace tools such as Docs, Drive, Gmail and Sheets
Large-context document and file work on supported models and plans
Web-backed answers and Google account integration, with features varying by country, account and subscription

How it works: The core is still a language model generating text from the context it receives. Its practical difference is the plumbing around that model: Search or a Google file can be retrieved and supplied as context, so you may not need to copy and paste it. Permissions and product settings decide what it can access; being in your Drive does not mean every file is automatically available.
What it costs: A free tier is available in supported countries. Paid access is commonly bundled through Google AI or Google One plans at about US$20 per month, with local pricing and Workspace arrangements varying.
Why this one: Choose it if your work already happens in Google Docs, Gmail, Drive or Sheets. Removing the copy-and-paste step can be worth more than a small difference in writing quality.
If you cannot get it: ChatGPT or Claude can work from files you upload. If a Google integration is not available on your account, use the assistant that is easiest to reach and control.

A fair test in one sitting

Choose one job that represents your week. Use the same kind of task you built on Day 4: turn meeting notes into an action list, draft a customer update or extract decisions from a report. Do not use a toy question that none of the tools will see again.

Prepare one clean input and one prompt. Remove private information you do not need to share. Do not improve the prompt for the second or third assistant. A fair comparison changes the tool, not the task.

Use the same prompt in each tool:
You are a [ROLE].
Turn the material below into [OUTPUT].
Use only the supplied material.
Format it as [CHECKABLE FORMAT].
Mark anything missing or uncertain. Do not invent names, dates or numbers.

Run the prompt in ChatGPT, Claude and Gemini using free access if possible. Save the three outputs with the tool name and date. If one tool reaches a limit or lacks a feature, record that as a result rather than quietly changing the test.
Score each output on accuracy, format, usefulness and the amount of editing you needed. Read for omissions and invented details, not just for pleasant wording. A confident error is a serious mark against a tool for that job.
Ask each assistant the same limit question. Compare whether it gives a specific answer about what it struggles with or simply advertises itself. This is not a perfect honesty test, but it shows how each one handles a boundary.
```
What can you not do well for this kind of work?
What should I check before I use your output?
```
Choose one default and write the exception. For example: use Gemini for work already in Google Docs, but use Claude for long policy reviews. Stop comparing until a real failure gives you a reason to retest.

See it

Take the same set of ten meeting notes to all three assistants. Ask for a decision list with three headings: Decision, Owner and Open question. Tell each one to use only the notes and to mark an absent owner as “not named”. That last rule gives you something concrete to inspect.

ChatGPT may give you a polished result and offer extra ways to format it. Claude may preserve a qualification that the others compress and may be more likely to say that no owner was named. Gemini may be the easiest if the notes already live in Google Docs, because the copying step disappears. None of those outcomes makes one assistant the winner for every person. They tell you what each shape feels like on this job.

Now look past tone. Did each assistant keep all three decisions? Did any turn a suggestion into a decision? Did any invent a date or assign the meeting organiser as the owner? How much editing did you perform before the result could be sent? A friendly answer that requires ten corrections is not faster than a plain answer that needs two.

The score may be close. That is normal. The difference between a four and a five is often less important than the difference between using a clear template and pasting a vague request. The purpose of the test is not to manufacture a dramatic verdict. It is to give you a defensible default and a reason to stop searching for a mythical perfect assistant.

Count the errors yourself

You cannot be told which tool fits your work. You have to measure it on a subject and task you understand, then choose the one with the best practical result.

Pick a real task from Day 4 and a topic you genuinely know well. You must be able to judge whether a name, date, process or claim is accurate without trusting the assistant to check it for you.
Use one identical prompt and one identical input in ChatGPT, Claude and Gemini. Ask for a structured answer and tell each one to mark missing information instead of guessing. Save all three outputs before you start scoring.
```
Give me five important claims about [your topic], each with its source and year.
Use a table with Claim, Source, Year and Confidence.
If you cannot verify a claim, write “not verified”.
Do not invent a source.
```
Check every claim in each output. Mark it verified, partly supported, cannot find it or does not exist. Count the errors and the unverified claims separately. A tool that admits it cannot verify something has behaved better than one that supplies a neat fiction.

Score the three tools out of five using the same sheet. Include accuracy, matched format, useful detail, ease of access and would reuse. Add one line about editing time. Do not score based only on which answer sounds nicest.

                         ChatGPT  Claude  Gemini
Accurate and checked?       ___/5    ___/5   ___/5
Matched the format?         ___/5    ___/5   ___/5
Marked missing information? ___/5    ___/5   ___/5
Editing time acceptable?    ___/5    ___/5   ___/5
Would reuse for this job?   ___/5    ___/5   ___/5

Ask each one the same two limit questions from the walkthrough. Record the answer, but trust your observed errors more than its self-description. A tool saying “I can make mistakes” is not useful unless it helps you find them.
Choose one default tool and one narrow exception, then save the decision. Write: “For my regular work I will use ____ because ____. I will use ____ only when ____.” Close the comparison tabs. You can retest after a real failure, not because another ranking appeared online.

What you should notice

Almost everyone finds that the scores are close, usually within a point or two. That is the useful finding. The differences between these tools are often smaller than the difference between a good prompt and a bad one, or between supplied source material and a question that asks the model to remember everything.

You may also find that “which model is smartest?” is not the question your week asks. The better questions are: which tool is reachable, which one handles your usual files, which one follows your format and which one makes its mistakes easiest to spot? A small integration advantage can save more time than a benchmark advantage you never notice.

The error count matters more than the prose. A fabricated statistic can read exactly like a true one: perfect grammar, a specific decimal and a source-shaped name. There is no texture that tells you it is false. That is why the tool's confidence or careful tone is not a verification method. Check the claim, especially numbers, names, dates, quotes and citations.

You also tested the safer pattern. Supplying links you trust turns an unreliable retrieval task into a more controlled reading task. The assistant can still misunderstand a passage, but you can point to the passage and correct it. That distinction tells you which jobs to hand over and which to keep.

So stop optimising the tool. You now have a tested, defensible reason for whichever assistant you keep as your default. Keep the exception if it is real, but do not turn every new release into a fresh identity crisis. Learn the one you use, keep your prompts, and measure the work that comes out.

A balance scale weighing three small equal weights on one side against one larger weight on the other. — Quality is only one weight. Access, file fit, limits and editing time belong on the same scale.

If you have five more minutes

Run the same scorecard again with a long document rather than a list of claims. Check which tool preserves a small qualification from page 17. That is a useful reason to keep a second tool available, but only if this is a task you actually do.

Tomorrow

Put the assistant to work on a document you have been avoiding.

Get the rest by email

One lesson each morning at 07:00 IST, with the task written out in full so you can do it from your inbox.

This is the course, not the newsletter. Joining does not subscribe you to the daily Brief.