A Tallinn studio workbench with a graphite keyboard, a pencil-ruled comparison notebook and a violet LED strip under the shelf.
Claude · ChatGPT Atlas · Gemini

Claude or Gemini: which fits your writing day?

General-purpose assistants look alike until you give them real work. Match the assistant to where your drafts already live — long writing, quick lookup, or multimodal handling.

Where each assistant lands

The fit board: six rows, one honest answer each.

Read down the left column to the job you actually have this week — a long report, a browser tab full of sources, a drive full of shared documents. The right column says plainly which assistant fits, and where it wobbles. Nothing here is a ranking; it is a shortlist matched to workflow.

Claude

Drafting and analysis — messy outlines turned into structure

ChatGPT Atlas

Browser-side help — recap, simplify, draft without leaving the tab

Gemini

Inside Google work — docs, mail and drive stay in one place

Multimodal input

Text plus images, voice or files — a photo of notes, an attached PDF

Prompt basics

Audience, length, format, source text — no course required

Cost sense

Free tiers first; pay when the limits interrupt paid work

Claude for drafting and analysis

Anthropic's conversational assistant, usually picked when a draft has to be reasoned through rather than just generated.

Give it a messy outline and ask for structure first; then ask for a tighter second pass in the same thread. It holds long context well enough for reports, seminar drafts and long emailed threads, which is where most solo writers actually notice the difference. Extra plugins rarely carry as much weight as a steady second pair of eyes on the same document.

Try it on: report outline · email thread summary · argument check

A hand photographing a paper document with a phone, its screen showing a blurred capture of the page in progress.

Same material, three handles

Multimodal input without overthinking it.

Multimodal simply means text along with images, voice, or files in the same prompt. Take a photo of handwritten notes, attach a PDF, or dictate an idea while walking. All three assistants handle some mix of this — the honest test is your material, not theirs.

Scanned handouts

Lecture pages, printed contracts, photographed whiteboards — one prompt, one pass.

Screenshots

Interface fragments and charts read better as images than as guessed descriptions.

Short voice memos

Draft a paragraph by speaking, then tighten the text in the same thread.

Pick the assistant that reads your actual material cleanly, not the one with the longest feature list.

The four lines that carry most prompts

  1. 1

    Name the audience

    Who is reading this — a first-year student, a client, your own teammate. Audience sets tone faster than adjectives do.

  2. 2

    Give a length

    200 words, five bullets, one paragraph. A number stops the model from guessing at scale.

  3. 3

    Pick a format

    Bullets, a short paragraph, a table, an email reply. Formatting is where most first drafts are lost.

  4. 4

    Paste the source

    Then ask for one revision at a time. Example: "Audience: first-year students. Length: 200 words. Format: bullets. Source below."

A two-week test you can run in Tallinn

Pick one, then run it on real work.

Week one: use one assistant for three real tasks and note time saved and fixes needed. Week two: repeat with the second option on similar tasks. Compare finished output, not feature lists. One page of notes beats ten review videos.

Students at TalTech and freelancers near Telliskivi have run this exact loop to pick a daily driver.

Someone running a two-week tool test at their desk, a marked comparison sheet on the wall and two AI chat windows open on the laptop.
Two assistants, three real tasks each, and a wall sheet that stays honest about what actually worked. Keep the printed comparison pinned next to the laptop while you test — the second week is where the useful difference shows.

Three honest limits

Where each assistant struggles.

The wobbles are predictable once you know the shape of them. Treat every first answer as a draft, and keep the verification step non-negotiable for client work.

Claude

Context goes thin

Long chats drift, and citations may be missing. Claude can still produce confident filler when the source is thin — if the input is one paragraph, the output rarely invents real depth, only the sound of it.

ChatGPT Atlas

Paywalls and names

Browser helpers can misread paywalled pages and quietly replace a name or a date with a plausible one. Keep prompts short, and check names and dates against the original page before you ship a reply.

Gemini

Sharing settings

Gemini can mix personal and shared files if settings are loose. Convenience and privacy travel together here — review sharing before you hand over sensitive documents or someone else's material.

Tone flattens when nobody reads it

All three assistants can produce a serviceable paragraph that sounds like no one wrote it. The fix is editorial, not technical: add one concrete detail only you would know, then cut the sentence that reads like a template. That step is where a draft becomes yours, and it is the one step no assistant can do for you.

Cost and free-tier sense

Pay when the limit interrupts paid work.

Most everyday writing fits free tiers at first. The honest trigger to pay is that a limit stops work you were already being paid for — not that a comparison chart said one plan is bigger than another. If you do pay, pay for one assistant only, then revisit in two months with a fresh look at what you actually used.

For a small team, one payer keeps the subscription clean, shared prompt patterns keep the value visible, and a review of the second seat usually answers itself: if nobody noticed the extra seat, it was not doing work.

What to check before you upgrade

  • Which task stopped. A specific limit on a specific job is the clearest signal that a paid tier will pay back. A vague feeling that you should upgrade rarely does.

  • How often it happens. Once a month is a different decision from every afternoon of the week. Write down the frequency for two weeks before you decide.

  • Whether a free alternative covers it. A different assistant's free tier, a shorter prompt, or splitting a long document across two sessions sometimes removes the wall entirely.

  • Who is paying. On a small team, one payer and one shared list of working prompts beat several quiet subscriptions nobody reviews.

Questions we get asked

Two short answers.

If you are between two tools and the decision keeps sliding, one of these two usually settles it.

ChatGPT, Claude, Gemini, or Copilot for daily work?
For writing and analysis, start with Claude or ChatGPT. For Google-centred files, test Gemini. For code-adjacent help, test a coding assistant instead. Match the tool to your files and habits rather than to a feature count — the assistant you already open daily usually wins on that alone.
Do I need prompt engineering courses?
No. Clear audience, length, format, plus the source text covers most needs. Save the prompts that work for you and reuse them — a personal library of ten good prompts is worth more than a course playlist on a topic this fast-moving.

Still unsure which assistant fits your files? Write to Dmytro with a short note about the documents you work with, and we will point at a starting tool.

Pick one assistant this week, then judge it on your own work.

Two weeks of real tasks will tell you more than any comparison page — including this one. If you want a second opinion before you start, a short note about the files you work with is enough.

Ask which assistant fits your files See research tools with sources

Ümera tn 2-152, Lasnamäe, Tallinn · Mon–Fri, 9:00–18:00 · [email protected]