Choosing an AI tool

Most AI apps offer three kinds of tool, each a different way of using AI. The job decides which. In a chat, you type a message and the model, the part of the app that writes, replies. You carry out what the reply suggests. An agent lets the model carry out several steps on your files or the web, then report back. A coding tool is an agent that builds things by writing, running and fixing code.

Team names for your netball club suit a chat, and a folder of receipts sorted by month for tax calls for an agent. A web page for your market stall suits a coding tool.

The more a mistake would cost you, the more you check. AI helps you think a decision through, such as whether to take a job. The decision stays with you, or with a professional for money, health or the law, because the app cannot answer for it.

• Use a chat to ask, an agent for many file steps or a coding tool to build, because each kind needs different checks, even when one app offers all three.
• Ask what a mistake would cost, then for an agent work on a copy of your files and check what it did, because you see the result, not each step.
• If your app can run calculations on uploaded files, upload the spreadsheet, not typed numbers, because the app works the sums out from the data instead of guessing at them.

1. Pick one real job you meant to do with AI this week, and say in a line what sort it is: asking or drafting (a chat), many steps on files (an agent), or building something (a coding tool).
2. Tell your AI app the job, and ask which of the three kinds of tool suits it and why. Ask too what a mistake on it would cost you and what you should check.
3. Do the job with the kind of tool your app suggested, as far as your phone allows. If the job touches files, make a duplicate of them first and use the duplicates.
4. Finish something you can keep, such as the reply sent, the files sorted or a web page you have saved. Try a web page yourself before anyone else sees it. If your phone cannot do the job, such as sorting a big folder, save the steps your app laid out in a note.

Expect your app to answer along these lines.

• A small job gets "a chat is fine".
• A job on files, or a build, gets an agent or a coding tool, with advice to work on a copy and read what it did.
• A costly job also gets a short list of what to check.
• If your app offers only a chat, it will say what it can do and what you would do by hand. Then do those parts by hand. If it suggests a kind of tool you cannot find on your screen, use only what you can see.

• An agent can change files in ways you did not mean, and some agents keep working without asking at each step. Work on a copy, look in the agent's settings for a way to make it ask before each step, and read what it did.
• A slower setting, which makes the model take longer and work through a reply more thoroughly, can still be wrong. A reply that shows how it got its answer can be wrong too. Check the facts you will act on.
• Your words and files go to the company that runs the app, so a new app means a new company. Before trying another app, write a line in your phone notes for what never goes into an AI, such as "No passwords or clients' names".

Choose the kind of tool that suits each job. Take more care where a mistake would cost you more, and keep the decisions that are yours.

• How AI works
• What not to share
• Connectors and permissions

• which-tool-for-which-job-1 (Choosing, waypoint 1): I can choose the kind of AI tool for a job (a chat, an agent on my files, a coding tool) and say how much care it needs. Proves: given two real jobs of theirs of different shapes, they choose a different kind for each, give the shape of the job as the reason, and say what a mistake would cost and what they would check.

Before a new kind of job, they say what shape it is and what a wrong answer would cost, and choose the kind of tool and the care on that basis, instead of using one kind of tool for everything. "Which AI is best?" becomes "which kind for this?".

• Explains: the job's shape comes first, not an app's name or "the biggest one"; a chat is for asking and drafting, an agent for many steps on files, a coding tool for building; what a mistake would cost decides how much care, and for an agent that means working on a copy and reading what it did. AI helps you think a decision through, and the decision stays theirs or a professional's.
• Ask: "Pick two jobs from your week that are different shapes. Which kind of tool for each, why, and what would a mistake cost?" Shows it: a different kind for each, the shape of the job as the reason (not a brand), and for each a cost of a miss and one thing they would check.
• Ask: "A friend wants a quick reply to a text. Another wants forty invoices in a folder renamed by date. Which kind of tool for each, and what care?" Shows it: the text gets a chat (nothing lost if the tone is off, and they read it before sending); the invoices get an agent on files, on a copy, with a check of the agent's settings for asking before each step and a few renamed files checked by hand, since a plain chat cannot touch the folder. Spare problem for a retry, never reused in the same chat: "A person wants a simple booking page for their pottery class", which gets a coding tool (or a chat that writes the code, which they must then run and test), on a copy, opened and tried by the person before anyone else sees it.

Hints for the person's AI. Write the words for the person, in the way their journal says they like to learn; never read these lines to them.

• Experiment: run the Try it with them on a real job of theirs. Once the job is done, optionally ask the app which model it is and compare the reply with the label on screen. The two can disagree, and the screen is the one to trust.
• Challenge: for a week, before each AI job beyond a quick question, they write one line on the job, what a miss would cost and which kind of tool they used. They notice how many stay a chat.
• Quiz (only if their journal allows quizzes): ask them to name a job for a chat, one for an agent and one for a coding tool, and what a miss would cost on each.
• Cards: five, one each for a chat, an agent on your files, a coding tool, "what would a mistake cost me?" and "the decision is yours".
• Game, "Pile it": read out jobs from their own week and have them sort each into chat, agent, coding tool or "my call, with AI to think". Add a twist, such as a client's private details, to show a job can change pile.
• Explain: sort their last week of AI use with them, then show where they used one kind of tool for everything.
• Project (Rucksack only; the Rucksack is coming soon): a file in their Rucksack that lists which of their apps is for which kind of job and which calls they keep.

• Teach the question, not the picker. Names, plan limits and settings change within weeks. Say "apps differ", read the person's own app on the day, and use only the per-app facts in For your AI, with their dates. Never say one app or model is best. Not for everyday chat: paid plans add stronger models, higher limits and some agents. Do not push a plan; whether to pay is "usage-and-plans".
• Avoid both overcorrections. "The biggest model for everything" is slow, uses up the person's allowance and adds little on small jobs. "Free is fine for everything" is wrong for some free models, which are built to be fast and low cost rather than careful; Claude's free plan has Sonnet, a strong general model, so say "some". Free is fine for low-stakes jobs, not for medical or legal questions (Mollick, 23 July 2026). The same for agents: "for developers only" and "hand over your whole computer" are both wrong.
• Help the person find their agent's own setting for asking before each step, and still work on a copy. An agent does not always ask first: Claude's agent on Pro, Max, Ultra and Team defaults to Auto, which keeps working without stopping at each step, and Manual makes it ask (Enterprise is Manual). It asks before permanently deleting files (checked 23 October 2026). The help pages differ by app.
• Name a slower setting in a clause and move on: it is not a guarantee, so the person checks what they will act on. The how is "thinking-modes" and "check-by-what-is-at-stake".
• People do take personal decisions to AI to think them through, such as a falling-out with a friend, a job offer or how to word a hard message, and that is a normal use. Help them lay out the options and what each might cost, then leave the call with them. For money, health and legal matters, AI can explain and draft, a second source or a professional makes the call, and the AI never gives the verdict ("big-decisions" owns the how).
• Building software is "vibe-coding"'s ground. Here the coding tool is one of three kinds and one line on when to reach for it.
• The model's description of itself is not evidence of which model answered. The screen is.
• Which tool never overrides the person's boundary ("your-boundary"). If they try a new app, they take their boundary line with them.
• Daypack case of a new app: the Daypack lives in the one app it was set up in. If the person tries a second app for a job, their Daypack does not go with them, and their AI must not suggest pasting their context across, because copies go stale. An app with no connector can get a Daypack of its own. Rucksack case: their fleet file records which app is for which kind of job.
• Do not sell the Rucksack as the answer to "which tool". The Rucksack is coming soon.
• If the person's one app lacks a stronger setting, an agent or a coding side, say so plainly and do not send them elsewhere unless they ask.
• Per-app names and plan limits date quickly: re-check by 21 December 2026 (checked 23 October 2026; see For your AI).

• Q: Which AI should I use: ChatGPT, Claude or Gemini? A: Start from the job: a chat app to ask and draft, an agent for many steps on your files, or a coding tool to build something. For everyday low-stakes jobs, any good one will do.
• Q: Which AI model should I use? A: Start with your app's default. For plans, maths or anything costly if wrong, switch to a slower, more careful setting where your app has one.
• Q: Why can't I see the newest model everyone is talking about? A: Which versions you can use depends on your plan, how far the app has rolled a new one out, and, at work, what your organisation allows. Check what your plan includes.
• Q: Is it safe to let an AI work on my files or computer? A: Only if you set it to ask before each change, because some apps work on without asking by default. Work on a copy and read what it did.

Facts (checked 23 October 2026 unless dated otherwise; fact check research/2026-10-23-essentials/batch-4-facts.md, cited below as "batch 4"; research note research/2026-10-20-concepts/which-tool-for-which-job.md, cited as "the note"). Where they differ, batch 4 wins. Menus and names move; read the person's own app on the day.

The three kinds. The app is what you open, the model writes the reply, and a harness lets a model use tools and take multi-step actions (Mollick, 18 February 2026, https://www.oneusefulthing.org/p/a-guide-to-which-ai-to-use-in-the; the note section 1). Do not use the word harness with the person. Microsoft's training module says to choose a mode by outcome: Chat for an answer, Cowork for a finished multi-step result, Code for a tool (16 September 2026, https://learn.microsoft.com/en-us/training/modules/what-generative-ai/4-copilot-workflow-companion). Mollick splits chatbots, where many options are good enough for low stakes, from agentic systems that "give an AI a computer" (23 July 2026, https://www.oneusefulthing.org/p/an-opinionated-guide-to-which-ai-b22). "Pick Claude or ChatGPT, pay the $20" is one educator's bet and dates within months: do not repeat it.

Where each kind lives (batch 4 fact 1):

• ChatGPT: Chat; Work (the agent, limited on Free and Go, wider on Plus; https://chatgpt.com/pricing, 5 November 2026); Codex (coding) (https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex).
• Claude: chat and Cowork are "one Claude" with "no mode to choose" (Pro, Max and Ultra first, then Team in beta and Enterprise where enabled, 16 September 2026, https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude); Claude Code is the coding tool (Pro, Max, Ultra, Team and Enterprise, not Free; https://claude.com/pricing, 5 November 2026). A "Claude Duet" preview (21 October) was seen and not researched: check on the day.
• Gemini: Spark is the agent (Ultra, 18 and over, US beta, 19 May; https://gemini.google/release-notes/); Antigravity is the coding tool, in the Mac app.
• Copilot: Chat, Cowork, Code and Autopilot; Code and Autopilot need the Frontier preview (https://learn.microsoft.com/en-us/training/modules/what-generative-ai/4-copilot-workflow-companion).
• Local: Ollama and LM Studio run open models on your own computer, free for local use; Ollama's cloud models are billed by usage, per token or from credits bundled with a paid plan (batch 4 fact 6, https://ollama.com/pricing, https://lmstudio.ai/pricing). "local-models" owns it. So an app that merges chat and agent still leaves the person deciding what shape the job is and how much care it needs.

Asking before each step (batch 4 fact 5). Claude's agent default on Pro, Max, Ultra and Team is Auto, which "keeps working without stopping to ask about each step"; Manual makes it ask; Enterprise is Manual; it asks before permanently deleting files (https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude). For other apps' agents, check the agent's own settings; the help pages differ. Mollick's advice is to leave agents set to ask for approval first until you trust them (23 July 2026, link above).

Calculating from a spreadsheet (batch 4 facts 2 and 3). It needs a code tool, not web search. ChatGPT "writes and runs Python" on uploaded data; data analysis is limited on Free and included on Go, Plus and Pro, it cannot make external web requests, and it "may not reliably extract exact values from image-based tables" (https://help.openai.com/en/articles/8437071-data-analysis-with-chatgpt). Claude: code execution on every plan including Free, switched on at Settings, Capabilities (6 August 2026, https://support.claude.com/en/articles/12111783-create-and-edit-files-with-claude). Work or school Copilot has a Python code interpreter through its agents, for Microsoft 365 Copilot licence holders and Copilot Chat users without metered usage enabled (https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/code-interpreter, read 5 November 2026). The Gemini app and personal Copilot: UNCONFIRMED, so do not promise it there.

Not seeing the newest model (batch 4 fact 4). ChatGPT: GPT-6 is "rolling out to eligible paid accounts"; plan, workspace admin, app version and limits all matter (https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt). Claude Fable: paid plan, app version, and "organization hasn't enabled" (https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan). Gemini: Argon is on Pro and Ultra, not Plus (https://support.google.com/gemini/answer/18648722, 5 November 2026). Copilot: chat experiences "vary depending on your tenant configuration and license" (https://learn.microsoft.com/en-us/copilot/plan/copilot-overview, 5 November 2026).

Choosing a model, per app (batch 4 per-app lines; menus and names move):

• ChatGPT: Free and Go get a Luna model with Instant and Think; Plus adds Medium and High; Pro adds Extra High and Pro. "Think harder" does not switch the level on Plus or Pro (https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt). The Luna version is UNCONFIRMED (the models page says GPT-5.6, the release notes and pricing say GPT-6), so say only "a Luna model".
• Claude: Free has Haiku and Sonnet, and Sonnet is the suggested default ("if you're not sure", https://academy.claude.com/tutorials/choosing-the-right-claude-model); Pro adds Opus; Fable uses credits on Pro and standard Team seats and counts toward the weekly limit (up to half) on Max, Ultra and premium Team seats (https://claude.com/pricing and the Fable help page above, 5 November 2026); an Effort setting.
• Gemini: with no Google AI plan you get Auto or Flash-Lite; Plus adds Flash and Pro, and Pro and Ultra add Argon; thinking Low (the default), Medium or High (https://gemini.google/subscriptions/, https://support.google.com/gemini/answer/18648722). "See response details" shows the model that answered (the note section 1).
• Copilot: personal modes are Quick response, Think Deeper, Study and learn, Smart (GPT-5) and Search; work or school Chat has Auto, Quick response and Think deeper (https://support.microsoft.com/en-us/microsoft-365-copilot/conversation-modes-in-microsoft-copilot).
• Free plans differ: some give a fast, low-cost model, but Claude's free plan has Sonnet 5.5, described as a versatile default for analysis and multi-step workflows (batch 4, "wrong or out of date").

A slower setting and a reply that shows its reasoning are not guarantees: Anthropic planted a hint and Claude 3.7 Sonnet said it used it 25 percent of the time, DeepSeek R1 39 percent (3 April 2025, https://www.anthropic.com/research/reasoning-models-dont-say-think). "thinking-modes" owns the how.

Data and the company. What the person types and uploads goes to the company that runs the app, so a new app is a new company; the boundary line travels with the person ("your-boundary"). A tool on your own computer keeps your words there, while a cloud option inside the same tool does not (batch 4 fact 6).

Questions people ask, matching the Questions section: which AI to use and which model to use (both homed here), why the newest model is missing, and whether it is safe to let an AI work on files. The last two come from the note section 6 and were fact-checked in batch 4 (facts 4 and 5). "Do I need to pay" is "usage-and-plans", "what are Instant and Thinking" is "thinking-modes", and "can I run AI on my own computer" is "local-models".

An example of this taught well, in passing, when someone is about to hand a folder to an agent: "That one changes your files, so work on a copy, check whether it asks before each step, and read what it did. Want the two-minute version of how to choose?"