You can talk to an AI in ordinary language. You do not need to learn code or memorize a library of “secret” prompts. The skill is closer to briefing a capable collaborator: explain the situation, say what you need, show what good looks like, and inspect the result.
Start with five useful lines
Say what you are doing
“I am preparing a five-minute update for a non-technical team.” Context changes the answer.
Name the outcome
Ask for the thing you need: an explanation, decision memo, outline, table, email, lesson, plan, or critique.
Give the relevant material
Paste the notes, attach the document, describe the audience, or link the source material when the product supports it.
Add constraints
Specify length, tone, deadline, format, must-include facts, and anything the AI should avoid assuming.
Ask for a check
For factual work, request sources, uncertainty, counter-evidence, or a separate verification pass.
A reusable prompt formula
Context: I am [situation]. Goal: Help me [specific outcome]. Audience: This is for [reader or user]. Material: Use [notes, file, link, facts]. Constraints: Keep it [length, tone, format]. Quality check: Before finishing, [verify claims / list assumptions / critique the result].
This is a scaffold, not a law. A simple question can still be one sentence. Add structure when the cost of misunderstanding is high or when the output needs to be reused.
Examples that produce useful work
Learn a difficult idea
Instead of “Explain compound interest,” try: “Explain compound interest to a 15-year-old using one worked example. Then ask me two questions to check whether I understood it.” The second version defines the learner, method, and feedback loop.
Improve writing without losing your voice
“Edit this email for clarity and remove repetition. Keep my direct tone, do not add claims I did not make, and show the three most important changes after the rewrite.” This preserves authorship while making the revision inspectable.
Research a decision
“Compare these three tools for a two-person design studio. Use current official pricing and documentation, separate facts from your judgment, and cite every time-sensitive claim.” That instruction makes the evidence standard explicit.
The first answer is a draft
Strong AI use is conversational. You can say “make the second paragraph less formal,” “show the assumptions behind that number,” “give me the strongest argument against your recommendation,” or “turn this into a checklist.” The follow-up should point to a specific weakness or next step.
“That is wrong” gives little direction. “The audience already knows the background, so remove the introduction and spend the space on implementation risks” teaches the conversation what to optimize.
Talk with voice, images, and files
Modern AI chat is broader than typing. Voice is useful when you need to think out loud, rehearse, or work hands-free. Images are useful for reading a chart, critiquing a layout, or identifying what needs closer inspection. Files let the answer stay grounded in your actual notes, policy, dataset, or draft.
Tell the AI how to use an attachment. “Read this PDF” is vague. “Extract the five eligibility rules from this PDF, quote the section headings, and flag anything that depends on a date” creates a checkable task.
Privacy and accuracy
- Do not paste passwords, payment details, private keys, or information you are not authorized to share.
- Remove unnecessary personal or client information before uploading documents.
- Verify medical, legal, financial, safety, and other high-stakes answers with qualified sources or professionals.
- Open citations. A link is not proof that the sentence is supported.
- Expect current facts—prices, laws, product features, schedules—to require a live search.
Doing it in AimiChat
AimiChat lets a conversation change modes without losing the objective. Start in regular chat, move a hard decision into Reasoning, research current evidence with Deep Research, inspect factual claims with AimiVerify, or listen to the answer with Voice.
The useful habit is to choose a mode because of the task's risk and output—not because a label sounds impressive. Everyday rewriting can stay fast. A consequential research report deserves sources and verification.
Five conversation patterns that reliably improve the result
1. Ask for questions before answers
Use this when the AI cannot know enough from one message: career decisions, business plans, health preparation, travel planning, or writing for an unfamiliar audience. Say: “Before answering, ask up to five questions that would materially change your recommendation. If a question will not change it, skip it.” The final sentence prevents an endless intake interview.
2. Separate transformation from invention
When editing your work, state which facts are fixed. For example: “Rewrite for clarity, but do not add dates, numbers, customer claims, or promises that are not in my source text. Put any suggested new claim under a separate ‘needs approval’ heading.” This is especially useful for marketing, reports, and client email.
3. Give a quality rubric
“Make it good” is not a test. Define three to five criteria. A useful brief might say the answer must be understandable without background knowledge, cite current official sources, name one serious counterargument, stay under 700 words, and finish with a recommendation that follows from the evidence. Ask the AI to score its draft against that rubric and revise once.
4. Request alternatives before committing
For a decision, ask for three materially different approaches—not three cosmetic rewrites. Require a low-risk option, a fast option, and a high-upside option; then ask what assumption would make each fail. This exposes tradeoffs before the model locks into one persuasive story.
5. Use an evidence boundary
When the answer must stay inside an attachment, say: “Use only the attached document. For every answer, cite the section heading or page. If the document does not establish something, say ‘not found in the provided material’ rather than using general knowledge.” Then manually inspect important references.
How to repair a weak AI answer
| Problem | What to say next | Why it works |
|---|---|---|
| Too generic | “Name the three decisions this answer is supposed to help me make. Rewrite around those decisions.” | Forces usefulness instead of topic coverage. |
| Too long | “Preserve the evidence and recommendation. Remove background the audience already knows.” | Cuts redundancy without deleting substance. |
| Confident but unsupported | “Extract each factual claim into a table with source, date, supporting passage, and confidence.” | Turns prose back into auditable units. |
| Wrong tone | Provide a 100-word example and say, “Match its sentence length, directness, and level of formality; do not copy phrases.” | A concrete example carries more signal than adjectives. |
| Misunderstood the goal | “Restate the audience, outcome, and constraints in your own words. Wait for confirmation before revising.” | Finds the broken requirement before another draft. |
A 30-minute practice session
- Choose a real task you can judge: an email, a one-page recommendation, a concept you know, or a plan for this week.
- Spend three minutes writing context, outcome, material, constraints, and a quality check.
- Generate a first answer. Mark every place that is generic, assumed, unsupported, or hard to use.
- Use one repair prompt from the table. Do not start a new chat; make the conversation carry the correction.
- Ask for a final self-audit against your criteria. Then verify one factual claim yourself.
- Save the prompt only if the second result required less correction. A long prompt that still produces rework is not a good template.
You should be able to explain why the final answer is better: it uses your material, serves a named audience, meets a visible constraint, and makes uncertainty inspectable. “It sounds smarter” is not enough.
Research behind this guide
OpenAI's current prompting fundamentals emphasize clear instructions, task context, iteration, and treating prompting as a conversation. Anthropic's official prompting guidance similarly recommends direct instructions, examples, roles when useful, and clear structure. The privacy and verification cautions follow the NIST Generative AI Profile, which treats confident false content and fabricated citations as a core risk, and the FTC's warning about sensitive or confidential information shared with model providers.