How to Use AI: A Practical Beginner's Guide

You do not need to learn programming or collect hundreds of “magic prompts” to use AI well. Start with one small task, explain what a good result looks like, and check the answer before you use it.

This beginner’s guide from Bakry Dev Hub gives you a repeatable workflow for ChatGPT, Gemini, Claude, DeepSeek, Grok, and similar assistants. Product buttons and model names change, but the method remains useful.

What can AI help you do?

General AI assistants are useful when the task involves language, patterns, or transforming information. Good beginner tasks include:

• Summarizing text or a document you provide.
• Rewriting an email for a specific audience and tone.
• Explaining an unfamiliar idea at your level.
• Turning notes into an outline, checklist, or table.
• Brainstorming options and criteria—not making the final decision for you.
• Extracting named fields from text.
• Reviewing code, explaining an error, or proposing tests.
• Comparing products when the assistant can browse and cite current sources.

AI is less reliable when you ask it to guess missing facts, calculate without verification, make a high-stakes decision alone, or answer a current question without access to current sources.

The safest mental model is: AI produces a candidate answer quickly; you remain responsible for whether that answer is true and appropriate.

Step 1: choose one clear, reversible task

Do not begin with “run my business” or “build my entire application.” Choose an output you can inspect and correct.

Good first tasks:

• “Turn these meeting notes into actions with an owner and due date.”
• “Explain this error message, then give me three checks to run.”
• “Rewrite this email to be concise and polite without changing the facts.”
• “Create a seven-day study plan from this course outline.”

Keep the original source and do not let the first AI answer overwrite important work. A reversible task lets you learn without making one model response a point of failure.

Step 2: choose the tool by input and outcome

You do not need the “strongest” model for every task. Ask these questions:

1. What will I provide? Plain text, a PDF, image, spreadsheet, audio file, code repository, or web question?
2. Does the answer need current information? If yes, use a mode that can browse and provide sources.
3. Does it need to take an action? Sending email, editing files, or operating software needs a tool-enabled assistant and careful approval boundaries.
4. Is the data sensitive? If yes, stop until you know the approved product and its data policy.
5. What format do I need? Paragraphs, bullets, JSON, a table, code patch, or step-by-step instructions?

For a current shortlist, see Best AI Models in 2026: ChatGPT, Gemini, Claude, DeepSeek and Grok. Beginners usually learn faster by choosing one accessible assistant for a week instead of changing models after every imperfect response.

Step 3: write a useful prompt

A prompt is simply the request and information you send to the assistant. Use this five-part formula:

Goal: What should be accomplished?
Context: What does the assistant need to know or read?
Constraints: What must it include, avoid, or preserve?
Output: What exact form should the answer take?
Verification: How should uncertainty, sources, or checks be handled?

You do not need every label for a simple request. The formula helps when the first answer would otherwise depend on guessing.

Weak prompt

Write an email about the delay.

Better prompt

Write a concise email to a client whose website launch will be delayed by two days.

Context:
- The payment integration needs one more security test.
- The new launch date is Thursday, August 6.
- We do not expect a price change.

Constraints:
- Be direct and calm.
- Do not blame the payment provider.
- Do not invent any compensation or guarantee.

Output:
- Subject line
- Email body under 140 words
- One clear next step

Before answering, list any missing fact that prevents an accurate draft.

The second prompt is better because the assistant does less guessing. It knows the audience, facts, tone, prohibited claims, format, and uncertainty rule.

Step 4: give source material, not vague memory tests

If you want a summary, provide the document. If you want a rewrite, paste the original. If you want code help, include the error, relevant code, expected behavior, and what you already tried.

Use delimiters so instructions and source material are easy to distinguish:

Summarize the source between <source> tags in five bullets.
Preserve every date and amount exactly.
Do not add facts from outside the source.
If a requested fact is missing, write "not stated."

<source>
[paste the text here]
</source>

This does not guarantee a perfect answer, but it makes omissions and invented details easier to detect.

For a large file, first ask the assistant to describe what it can access: file name, page count or sheets, visible headings, and any unreadable sections. That catches failed uploads and incomplete parsing before you rely on the output.

Step 5: turn the first answer into a short review loop

The first response is a draft. Review it against the goal and then ask for a targeted revision.

A useful three-turn loop is:

1. Draft: “Create the answer using these constraints.”
2. Critique: “Check the draft against each constraint. List unsupported claims, missing information, and ambiguous wording.”
3. Revise: “Produce a corrected final version. Do not add facts that are not in the source.”

Do not ask only, “Are you sure?” The same model may confidently defend its earlier mistake. Point it back to the source, acceptance criteria, calculation, test, or official documentation.

If the result still fails after two focused revisions, change the context, split the task into smaller pieces, or test another model. Repeating the same vague prompt is unlikely to solve a missing-information problem.

Step 6: verify before you copy, send, or publish

Verification depends on the task.

Facts and research

• Require links to primary sources for current claims.
• Open every important citation and confirm it supports the sentence.
• Check publication date and whether the page describes the same product or version.
• Separate facts from the assistant’s inference.
• Search independently for a high-impact claim.

An AI-generated link can be wrong or can point to a real page that does not support the claim. A citation is evidence only after you inspect it.

Numbers and spreadsheets

• Recalculate important totals with a calculator or spreadsheet formula.
• Confirm units, currencies, tax treatment, date ranges, and rounding.
• Ask the assistant to show the formula, then verify it independently.
• Test edge cases such as zero, missing values, negative numbers, and duplicates.

Code

• Read the diff; do not paste changes blindly.
• Run formatting, type checks, tests, and a production build where relevant.
• Check security boundaries, error handling, dependencies, and destructive operations.
• Ask what assumptions the fix makes and what remains untested.

Code that looks plausible can still call a nonexistent API or create a subtle regression. Executable verification is stronger than a confident explanation.

Writing and communication

• Confirm every name, date, amount, promise, and quotation.
• Remove private information that is unnecessary for the reader.
• Read it in your own voice.
• Make sure the message does not claim experience or evidence you do not have.

Step 7: protect private and sensitive data

Before you upload text, a file, screenshot, or repository, ask: Am I allowed to share this with this product under this account?

Do not paste:

• Passwords, API keys, private tokens, or recovery codes.
• Customer records, identification documents, or medical information.
• Confidential contracts, unreleased financial data, or private source code without approval.
• Production logs that contain session IDs, email addresses, IP addresses, or secrets.

Redact data or replace it with representative examples. Check the current provider settings and policy for consumer chat, business workspaces, and API use separately; they may not have the same retention or training controls.

For high-impact medical, legal, financial, employment, or security decisions, use AI to organize questions and information—not as the final authority. Involve a qualified person who can take responsibility for the decision.

Practical prompts for common beginner tasks

Learn a subject

Teach me [topic] at beginner level.
First ask three questions to understand what I already know and why I need it.
Then give me a 20-minute lesson, one worked example, three practice questions,
and an answer key hidden under a separate heading.
Flag any part that depends on a current version or law.

Use the practice questions before reading the key. Otherwise the assistant becomes a summarizer instead of a tutor.

Summarize a document

Summarize only the attached document for a manager who has five minutes.
Return: the decision, five supporting facts, unresolved questions, and next actions.
Cite the page or section for every fact. If the file does not state something, say so.

Compare options

Help me compare [option A] and [option B] for [specific use case].
First propose five decision criteria and ask me to approve or edit them.
Then use current primary sources, link each source, separate facts from inference,
and recommend an option only after applying the agreed criteria.

Debug code

Explain this error before proposing a change.
Expected behavior: [describe it]
Actual behavior: [describe it]
Environment and versions: [list them]
What I tried: [list it]

Return:
1. The three most likely causes in order.
2. One check that confirms or rejects each cause.
3. The smallest safe fix after the cause is confirmed.
4. A test that would prevent regression.

Do not invent files, APIs, or command output you have not seen.

A 20-minute first AI session

Use a real but non-sensitive piece of text, such as public documentation or your own rough notes.

1. Minutes 0–3: define one output and who will use it.
2. Minutes 3–6: write the prompt with goal, context, constraints, output, and verification.
3. Minutes 6–10: run it and mark every sentence that contains a fact.
4. Minutes 10–14: verify facts against the source and list problems.
5. Minutes 14–17: ask for one revision using that list.
6. Minutes 17–20: save the final prompt, the corrected result, and one lesson for next time.

Your success measure is not “the AI answered.” It is “I produced a result I can inspect, verify, and safely use.”

Common beginner mistakes

• Asking for a broad outcome without an audience or success criteria.
• Treating polished language as proof that facts are correct.
• Asking the model to summarize a document it cannot actually access.
• Using web research without opening the citations.
• Uploading sensitive data before checking policy and approval.
• Letting AI make an irreversible or high-impact decision alone.
• Changing models repeatedly instead of improving context and verification.
• Copying code without running tests.
• Publishing generic AI text that adds no experience, evidence, or useful decision.

Avoiding these mistakes matters more than memorizing elaborate prompt tricks.

Frequently asked questions

Do I need coding skills to use AI?

No. General assistants accept natural-language requests. Coding becomes relevant when you want to automate an API workflow or build AI into a product.

How do I write a good AI prompt?

State the goal, relevant context, constraints, desired output, and how uncertainty should be handled. Provide the source material and an example when format or tone must be consistent.

Can I trust AI answers?

Treat them as drafts. Trust should come from verification: primary sources for facts, calculations for numbers, tests for code, and qualified review for high-stakes decisions.

Can AI read a PDF, image, or spreadsheet?

Many assistants can, but support depends on product, plan, file type, and size. After uploading, ask what the assistant can actually see before requesting analysis.

Is AI free to use?

Many products offer free access with changing limits. Paid assistant plans and developer API billing are separate. Check the current plan and pricing pages for the product you intend to use.

Should I say “please” or use special prompt words?

Politeness is fine, but clarity has more effect. Explain the task, source, constraints, output, and verification. There is no universal magic phrase that replaces good context.

Official learning references

• OpenAI prompting guide
• Google prompt design strategies
• Anthropic prompting best practices
• NIST Generative AI Profile

Have a question about this guide or an idea for a technical collaboration? Contact Bakry through the Dev Hub.