---
title: "How to Use AI: The Complete Practical Guide"
description: "A practical, substantial guide to using AI for thinking, research, writing, coding, local files and computer use—with projects, prompts and a first-week plan."
author: Power of AI
date: 2026-09-06
url: https://powerofai.ca/guide/how-to-use-ai
---

# How to use AI: from your first question to things you never thought you could build

A practical, substantial guide to using AI for thinking, research, writing, coding, local files and computer use—with projects, prompts and a first-week plan.

Give your curiosity somewhere bigger to go.

There is a point where using AI changes character. You stop asking for an explanation of something and start making something of your own: a searchable collection, a working calculator, a researched essay, a map you can explore. The conversation becomes a way into work that used to feel out of reach. This guide is about getting to that point—and learning enough to know whether the result deserves to be used.

This builds on Graham’s original [Be curious, try everything](/article/be-curious-try-everything) article. Ethan Mollick’s [practical AI guides](https://www.oneusefulthing.org/p/using-ai-right-now-a-quick-guide) are recommended further reading. The chapters and exercises here are original to Power of AI.

## 1. Start with the thing you keep putting off

Forget the perfect prompt for a moment. Think of a real piece of unfinished work. Perhaps you have a folder of notes you cannot search, a spreadsheet you do not quite trust, or an idea for a little website. Choose something you understand well enough to judge, even if you do not know how to produce it. That combination gives you both motivation and a way to spot nonsense.

Make the first session small enough to finish. A searchable page containing five public notes is a better first project than an entire knowledge management platform. You get to experience the whole loop: explain the need, inspect a result, change it, and use it. The point of a small first version is to create something you can react to. It can become ambitious very quickly once the basic idea works.

Your job is to keep supplying reality. Who will use this? What information is missing? Which sentence sounds nothing like you? What happens when somebody enters zero? The assistant can produce many candidates; you decide which ones fit the world you are working in.

### A request you can adapt

Help me turn [unfinished task] into something useful today. I have [materials] and understand [topic], but I do not know [technical part]. Propose a small first result I can inspect in one sitting, then help me make it. Use a sample if the real inputs are missing. Label the sample clearly.

## 2. Choose a working environment, then a model

The question “Which AI is best?” leaves out the most useful detail: best at doing what, with what access? A chat box, a source notebook, a coding agent and an assistant operating a browser are different working environments. Even a capable model cannot inspect a file it has never received or click a button without the appropriate tool.

For a conversation, explanation or first draft, start with a chat you can already use. For a question about several documents, choose an environment that can accept those documents and let you inspect the evidence. For a website or repeatable data transformation, use an environment that can create files and run code. For an interaction that exists only inside an app, use supported browser or computer tools.

Ask the assistant what it can actually access in this task. Have it name the files it opened or the tool it used. An answer that describes how you could analyze a spreadsheet is different from a completed analysis of your spreadsheet. Make that distinction early and you avoid spending twenty messages polishing an imaginary result.

Try your existing access before buying another subscription. If a limitation repeatedly blocks a task you care about, identify that limitation precisely: file handling, usage allowance, image generation, or a missing integration. Check the provider’s current terms and access details. A list of fashionable model names is a poor substitute for knowing what you need.

[Find the right kind of AI tool](/finder)

## 3. Learn the second question

A first response tells you what the assistant understood. Treat it as the beginning of the collaboration. “Be better” rarely explains what needs to change. “The audience already knows the background; lead with the decision and keep the evidence beneath it” gives a clear direction. Point to the paragraph, example or behavior you want changed.

When you are uncertain yourself, make the conversation investigative. Ask for two plausible approaches and the tradeoff between them. Ask which missing fact would most change the answer. Supply that fact, then ask for a revised result. You are building an understanding together rather than hoping your opening message contains every useful detail.

Use examples to communicate taste. If you want an article with your voice, supply a short passage you wrote and explain what you like about it: direct opening, concrete details, occasional humor. Ask the assistant to preserve your position and flag any new claim. Examples help with style; they do not give permission to invent your experiences.

### A request you can adapt

Here is what works in this draft: [specific detail]. Here is what misses: [specific detail]. Revise only the affected parts. Keep [facts or wording] intact. Explain any new assumptions you introduced in a short note after the draft.

## 4. Use AI to make your thinking visible

Some of the best uses produce a better question. Give the assistant a rough argument and ask it to identify the assumption holding the argument together. Ask what observation would weaken your conclusion. Turn a vague disagreement into two competing explanations, then list the evidence needed to distinguish them.

For learning, make yourself answer before the assistant explains. Try a short diagnostic question, work through your answer, and request feedback tied to your mistake. A fluent explanation can feel like understanding. Solving a fresh example without the chat open is a much stronger signal that you learned something.

For a difficult conversation, rehearse one exchange at a time. Ask for a reasonable person with a different view, rather than an opponent designed to lose. Afterward, inspect where you dodged a question or used a phrase that could sound accusatory. The rehearsal cannot predict another person, but it can help you choose your words.

[Learn without outsourcing the learning](/learn/learn-with-ai-without-outsourcing)

## 5. Turn research into a trail someone else can follow

A long answer with links is a starting point. A useful research artifact tells you what was asked, which sources were examined, what each supports, and what remains unsettled. Request a source table alongside the narrative. Give every consequential claim a location you can revisit: a page number, section heading, row identifier or direct source link.

Suppose you are researching the history of a local building. A modern tourism page, a newspaper article and an original survey might all discuss it. They are doing different jobs. Keep the source date separate from the date of the event. Keep a recorded location separate from a proposed identification. Ask the assistant to preserve those distinctions before writing the story.

When sources conflict, ask what evidence could resolve the conflict. Do not ask the model to average two incompatible claims into a smooth paragraph. It may be appropriate to finish with an unresolved question. That is useful research if the next person knows exactly what to look for.

Build a reusable collection as you go. A small local page with source titles, excerpts, dates and filters can be easier to work with than a growing chat. You can ask a coding agent to make that page from approved materials. The research and the software begin to reinforce each other.

### A request you can adapt

Research [question] using [permitted sources]. For each finding, give the claim, source, source date, exact supporting location and limitation. Separate primary records from later accounts. Keep contradictions visible. Create a short narrative only after the evidence table is ready.

[Build a research collection you can explore](/workflows/research-to-website)

## 6. Give creative work a direction worth exploring

“Make it professional” tends to remove interesting choices. Describe what the work should do to its reader or viewer. A neighborhood event poster might need to feel lively and approachable while remaining readable from across a room. A personal essay may need a blunt first sentence and room for uncertainty. Those are useful creative constraints.

Ask for three substantially different directions before asking for dozens of minor variations. For a story, change the source of tension. For a website, change the information hierarchy. For an image, change composition and lighting. Select a direction, explain why, and develop it. Keep a record of the choices you want preserved so each revision does not start from scratch.

You can also create tools for creative work: a character timeline, a searchable reference board, a rhythm practice app, or a way to compare headlines without seeing who wrote them. AI becomes more interesting when it helps you design the workshop as well as the work inside it.

[Use the writing workshop](/writing)

## 7. Coding is how you give an idea behavior

A website can respond to a choice. A script can repeat an operation across a hundred files. A little calculator can let you explore a decision. You do not need to start by understanding every part of a software stack. Start by describing the behavior: what a person supplies, what the tool returns, and how you will tell whether it is right.

For a first build, choose a tool that can run locally using sample data. A reading tracker, practice quiz or CSV checker gives you useful interactions without immediately introducing accounts, payments and databases. Ask the agent to implement one complete path through the experience and show it running. A convincing screenshot alone does not establish that the controls work.

Then become its most awkward user. Leave a field blank. Enter zero. Paste a long title. Refresh after saving something. Open it on a narrow screen. Ask the agent to reproduce each problem and fix the cause. This is a skill you can develop immediately: describing a defect clearly and verifying that the behavior changed.

Learning some code makes this collaboration more powerful. Ask which function handles an action, what data it expects, and why a particular check exists. Make a small edit yourself. You can build and learn in the same session, with each working feature giving the technical explanation a reason to matter.

[Build your first useful tool with Codex](/codex/build-your-first-tool)

## 8. A folder of files can become a workspace

A local project changes the scale of a task. Instead of pasting one excerpt, you can work with a collection of approved files and produce several connected outputs: an inventory, a cleaned table, charts, a report, and the script needed to repeat the process. OpenAI documents local folder projects separately from projects based on uploaded or connected sources. File access still depends on the environment and its permissions.

Create an input folder and an output folder. Put a few copies of suitable materials in the input folder first. Ask the agent to inventory them before transforming anything. The inventory should say which files it actually read, which formats it could not parse, and where the important information lives. A filename is a clue, not the content of a document.

For a recurring task, ask for a repeatable transformation. “Clean this export” can turn into a script with a documented input format, an exceptions report and an example command. On the next export you can rerun the same operation and compare counts. That is a more durable gain than manually reconstructing the conversation every month.

Local file access does not mean the AI processing is necessarily offline. Check the service’s data handling and your organization’s rules before supplying sensitive material. Start with public or synthetic examples while you learn how the workflow behaves.

[Understand file access and build a useful folder workspace](/workflows/ai-file-access)

## 9. Let it operate the interface when the interface matters

Computer use means an assistant can interact with an allowed graphical app: inspect the screen, choose controls and enter information. Browser use is particularly useful for checking a website you are building. A dedicated integration is often a better choice for retrieving structured records. Choose the method that gives you a result you can verify.

For your first computer-use task, give it a harmless, observable workflow. Ask it to open a local page, try the search, clear the filters, resize the view and report what happened. For a desktop app, name the app and exact window. Explain the end state you expect. The assistant should inspect what changed after an action instead of treating a click as proof of success.

The official OpenAI setup currently supports Computer Use on Windows and macOS in supported regions, with the plugin enabled and appropriate app permissions. Windows uses the active desktop, so it needs the foreground while it operates. Availability and controls can differ by account or workspace. Follow the linked setup guide for your environment.

Once a workflow involves a real account, specify whether the result is a draft or an action. Preparing an email and sending it are different outcomes. Keep a clear stopping point at the reviewable result when you want to make the final decision yourself.

[Explore the computer-use field guide](/computer-use)

## 10. Work with Codex as a project collaborator

A productive Codex task has an outcome, a working location and a way to judge completion. “Fix the site” is hard to evaluate. “Make the search preserve its query on refresh, add an empty state, and verify both behaviors in the browser” is concrete. Give the agent enough freedom to inspect the project and choose an implementation while preserving the behavior you care about.

In an existing project, begin with orientation. Ask it to read the instructions, identify the entry points and check for existing changes. Then have it connect the requested behavior to the relevant code. This helps avoid an impressive rewrite that solves the wrong problem or erases someone else’s work.

A useful finish includes the changed artifact, how to run it, what was checked and what is still uncertain. You should be able to continue the work tomorrow. For a longer project, ask for a concise handoff describing the current state and the next concrete step. Keep repeated project expectations in an instruction file so they are available beyond one conversation.

[Go deeper in the Codex handbook](/codex)

## 11. Turn a good session into a repeatable advantage

After a workflow helps twice, look at what stayed the same. You probably repeated the input requirements, the output format, the checks or the style preferences. Put those stable parts in a short project document, template or skill. Keep this week’s data out of the general instructions.

Before scheduling the task, run it manually with an ordinary input and an awkward one. Try a missing file, a duplicate row or a source that has not been updated. Decide what the assistant should do in each case. A weekly report that explains a missing source is more useful than one that silently fills the gap with guesses.

Track the total work, including preparation and review. A draft that appears in two minutes may still require forty minutes of checking. Some projects are worth doing because they make a new capability possible, even if they save no time. Name the benefit you are actually pursuing: speed, consistency, understanding, creative range or access to work you could not do before.

[Make your workflow reusable](/codex/skills-and-repeatable-work)

## 12. A first week that takes you somewhere

Day one: choose one task you understand and make a usable result. Keep the first request and the final version. Write down which piece of context most improved it. Day two: ask the assistant to help you learn something, then solve a fresh example yourself. Day three: work with two sources and verify the important references.

Day four: bring a small set of approved files and produce an artifact you can open outside the chat. Day five: build a tiny interactive tool. Day six: test it with awkward inputs and ask for focused repairs. Day seven: package what worked into a brief, a README and a next-step list. These are practice sessions, not a promise that every project will finish on schedule.

Choose a thread that connects the week. If you care about local history, your question, source collection, file organization and first website can all serve the same project. If you coach a team, your learning exercise, spreadsheet and practice planner can connect. Familiar subject matter gives your experiments continuity.

Then raise the ambition. Add a map. Build an importer. Compare versions. Make the explanation interactive. The next step should be more interesting because you now have a working foundation and a better sense of what to check.

[Pick a project and try the working showcase](/showcase)

## Take this into your next task

Choose a real task. Give the assistant useful context and the tools it needs. Make something you can inspect. Push the result further through specific feedback. Keep the parts that work and use them to attempt a more interesting project.

[Build a project brief](/showcase#brief-builder).

## Questions people ask

### Do I need to know how to code to use AI well?

No. Start with work you can judge: a draft, a source comparison or a small tool with obvious inputs and outputs. Coding knowledge becomes useful as projects grow, and you can learn it while building.

### Can AI really access files on my computer?

A supported local agent can work with files made available to its task, subject to permissions. An ordinary web chat does not automatically see your folders. Use approved uploads, a connected source or a local project.

### What should I try after basic chat?

Create a finished artifact from a small set of real inputs, then check it outside the chat. A source table, spreadsheet or tiny local app makes the next level of capability tangible.

## Sources and editorial notes

Reviewed September 6, 2026. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.

- [OpenAI: prompting and context](https://learn.chatgpt.com/docs/prompting)
- [OpenAI: projects and local folders](https://learn.chatgpt.com/docs/projects)
- [OpenAI: Computer Use, setup and platform requirements](https://learn.chatgpt.com/docs/computer-use)

## Keep going

- [The Codex Handbook: Build, Debug and Work with Files](/codex)
- [AI Showcase: Working Tools and Project Blueprints](/showcase)
- [How AI File Access Works: Uploads and Local Folders](/workflows/ai-file-access)