I just want help with writing and questions

Dana dropped four screenshots into Slack at 6:12 p.m. Same thank-you email. Four chat boxes. ChatGPT had a warm first paragraph and a fake invoice number (INV-4419). Claude had the right names and a stiff close. Gemini had rewritten the Google Doc in place, including a sentence Dana had already deleted. Grok had a punchy line about “shipping the win” that nobody on the team would say out loud. The thread had 11 replies. Nobody had hit Send. Forty-seven minutes, four apps, one email still sitting in drafts.

This is Part 1 of Which AI product should I use? If you mostly want help writing and asking questions, you do not need a coding agent, a folder bot, or a local model. You need one chat box, a 20-minute habit, and a rule for when the other brands are actually shorter. Plan names and buttons move. Treat this as a dated map and re-check the product the week you roll it out.

What you’ll learn

What “writing and questions” includes, and what it does not (code folders, payroll files, image studios)
How ChatGPT, Claude, Gemini, and Grok differ when the artifact is a paragraph, not a diff
A try-this-first default you can say in one breath
A 20-minute first session, including a paste-ready brief
The mix-ups that burn 47 minutes and invent invoice numbers

The job is a paragraph, not a project

People say “I need AI” the way they used to say “I need Excel.” That sentence hides the object. Is the output a thank-you, a definition, a study recap, or a 12-file refactor? This part is only the first family of jobs: drafts, explanations, rewrites, and questions. The unit of work is a conversation. You type, you maybe attach a file you are allowed to share, you read, you push back, you copy something into an email. You do not grant folder permissions. You do not install a CLI.

If the work is “make this sound like Dana,” stay here. If the work is “edit the repo,” that is Part 2. If the work is “clean the shared drive,” that is Part 3. Mixing those jobs is how you get four screenshots and no Send.

Four writing jobs a chat box can take: draft, explain, question, and rewrite, each with a one-line teaching note

Four chat doors for the same job

As of writing, four consumer products will take a writing job without pretending to be your IDE. They are not interchangeable skins. They also are not magically different souls. The useful split is where the work already lives and what you will verify.

Product	Company	Open this chat if	Watch for
ChatGPT	OpenAI	You want a general writing partner and a huge share of “just ask”	Warm tone that invents specifics (Dana’s INV-4419)
Claude	Anthropic	You want a careful draft, a long paste, or a blunt “this claim is thin”	Stiff voice if you forget to name the audience
Gemini	Google	The file is already a Doc, Sheet, or Gmail thread	It will edit the living file if you let it. That is a feature and a foot-gun
Grok	xAI	You want a live-flavored pass, or you already live in grok.com / X	Punchy lines that are not citations. Open the links

Full brand tracks live on this site if you later want depth: ChatGPT, Claude, Gemini, Grok. This chooser is the hallway, not the whole building. Button labels, Free vs paid SKUs, and model nicknames change. Confirm on chatgpt.com, claude.ai, gemini.google.com, and grok.com the week you pick a default.

What “good at writing” actually means

Vendors will always have a new score. Ignore the leaderboard for this job. A writing chat is good if it (1) keeps your names, (2) does not invent numbers, (3) matches the voice you asked for after one correction, and (4) is the app you will actually open on a Tuesday. Dana’s ChatGPT draft felt human and still shipped a fake invoice. That is a verification miss, not a reason to collect three more subscriptions before lunch.

Rule of thumb: One chat box per job family until that box fails twice on the same kind of task. Then switch for that job, not for your whole life.

Try this first

Defaults are not trophies. They are a way to stop the screenshot thread. You can change them after two weeks of real work.

Try this first for writing: ChatGPT or Claude for a paragraph today, Gemini if the file is already a Google Doc, Grok for a live-flavored…

Default: ChatGPT or Claude in the browser. Pick the one whose first draft you already prefer after one honest try. If you have no preference, start with ChatGPT for everyday notes and Claude when the paste is long or the stakes are “this will be forwarded to finance.”
Already in Google: Gemini. Walking the same thank-you out of Docs, into another app, and back is how Gemini’s in-place edit looks like a bug. It is the point.
Need a now-ish web flavor: Grok chat, then open whatever it cites. A live tone is not a footnote.
Work account vs personal: If the email is for customers, use the account your company pays for. Part 3 covers files. The same rule applies to a lonely paragraph: payroll language does not belong in a free personal chat.

A 20-minute first session

Do this once with a real draft sitting on your desk. Dana’s thank-you is a fine toy if you swap in your own names.

Open one chat box. Close the other three tabs.
Paste the current draft, not a vibe. If you have no draft, write four ugly bullets first. The model is better at rewriting than at mind-reading.
State audience, length, and what must stay true. Names, numbers, dates, and promises are sacred. Tone is not.
Ask for one version, not five. Five versions is how 47 minutes happens.
Read it out loud. Kill any sentence you would not say. Check every digit against the source (order 1182, not INV-4419).
Send from your mail client. The chat box does not get to press Send.

A brief you can paste. Swap the bracketed bits.

You are helping me rewrite a short thank-you email.
Audience: Maya Chen, who runs ops at Northpark Dental.
Length: 120-150 words, one screen on a phone.
Must stay true: we shipped order 1182 on Tuesday; the backorder was our delay, not theirs; I am Dana in billing.
Voice: warm, specific, no slogans, no "circling back," no fake invoice numbers.
Here is my ugly draft:
[paste]

Return: (1) a subject line, (2) the email, (3) a 3-bullet list of facts you used so I can check them.

What that brief does: it names the person, the length, the sacred facts, and a check list. If the model returns INV-4419, you catch it in step (3) instead of in Slack at 6:12 p.m.

Worked example: Dana’s thank-you

Ugly draft, labeled as an example:

Field	Value
To	Maya Chen, Northpark Dental
Order	1182
What happened	Shipped Tuesday after our warehouse delay
Dana’s ugly line	“sorry again for the mess, boxes should be there, let us know if counts are weird”

A pass that would survive finance: thank Maya for waiting, name order 1182, own the delay in one clause, tell her who to ping if a line is short, skip the slogan. A pass that would not: a poetic opener, a fabricated credit memo, and “we’re thrilled to partner.” If Gemini already edited the Doc, Dana still has to diff it. In-place is faster. It is not a substitute for reading.

Questions use the same muscle. “Why is Tuesday’s revenue 12,403 when the dashboard said 11,880?” is a writing-and-questions job only if you want an explanation of possible causes. It becomes a data job the moment you need the warehouse table. For that split, see Practical AI for analytics people. Chat can list hypotheses. It cannot sign the number.

Common mistakes

Running the same prompt through four apps “to compare.” You are buying a committee. Compare after the first tool fails twice.
Pasting customer names into a personal free chat because the work laptop was at home. Use the work account or wait.
Asking for “make it better” with no audience and no draft. You will get brochure paste.
Trusting a number the model invented because the sentence sounded confident. Require a facts list. Open the source.
Installing Claude Code or Grok Build because a tweet said “agents write emails now.” That is a different door. Stay in the chat box until a folder is the actual job.
Letting Gemini rewrite the living Doc, then noticing an hour later that the deleted apology is back. Version the file or ask for a comment, not a silent replace, until you trust the habit.

Free vs paid, without a shopping spree

Free tiers exist on all four and are enough to learn the 20-minute session. Paid plans buy higher limits, longer context, and sometimes work-admin controls. They do not buy a personality transplant. If Dana hits a cap after three thank-yous, that is a limit, not a sign she needs a coding agent. Phase P’s setup series will walk Free vs paid without hype. For this page: start free or on the work plan you already have. Add a second paid seat only after 30 days in one box.

Mobile apps are the same job in a smaller window. They are not a new product family. Use them for questions on the train. Do not dictate customer names in a quiet car. That is Part 4. Desktop apps (ChatGPT, Claude, Grok) are still chat boxes unless you open the coding or cowork surface on purpose. The icon on the dock lies. Check which mode you are in.

How to practice this week

Pick one recurring writing job (standup note, parent email, vendor chase). Run the 20-minute session three times in the same chat box. If you still hate the voice, switch once. Then stop collecting logos. When you are ready to pick among jobs, not brands, Part 6 is the decision tree. If what you actually wanted was help on a codebase, skip ahead to Part 2. If you care about what “open” even means, that is a later series; Part 4 is the privacy fork.

Quick recap

Writing and questions live in a chat box. Not in a repo agent.
Try ChatGPT or Claude first. Gemini if the file is already Google. Grok if you will open the links.
One version, sacred facts, you press Send.
Four screenshots is a process failure, not research.

Sources

Product surfaces and plan names move. Confirm on the vendor site the week you adopt a default.

OpenAI: ChatGPT (consumer chat)
Anthropic: Claude (consumer chat)
Google: Gemini (app and Workspace-adjacent chat)
xAI: Grok (grok.com chat)
Analytics Made Simple: Learn (series index)
Practical AI for analytics people (when a number still needs a table)