How to Talk to AI
A practical copy and paste guide for using current frontier AI models without pretending they are magic.
Use this guide like a playbook. Start with the quick start. Learn what changed with frontier models. Pick the right model for the work. Then steal a template, paste it into your AI tool, and improve the result in rounds.
The 5 Minute Quick Start
With current frontier models, the best prompt is less about magic wording and more about managing the work. Tell the model what outcome you want, what success looks like, what context it must use, and where it should stop and ask questions.
You are my assistant for this task.
Outcome:
[What I want done]
Success criteria:
- [What a good answer must include]
- [What would make this useful to me]
Context:
[Paste notes, links, files, screenshots, constraints, audience, examples, or prior drafts]
Output:
- Format: [memo / checklist / table / plan / email / JSON / other]
- Length: [short / medium / detailed]
- Tone: [plainspoken / executive / friendly / direct / other]
Constraints:
- [Deadlines, budget, tools, policy, source limits, or things to avoid]
Evidence rules:
- If this depends on current facts, use current sources and link them.
- Separate confirmed facts from inference.
- Say when you are unsure.
Work style:
- Ask only the questions that materially affect the answer.
- If you have enough information, proceed.
- Give me the best first version, then list what you would check or improve next.What Changed With Frontier Models
The biggest change is that the best models are no longer just chat boxes. They can reason through longer work, use tools, read files and images, write code, browse when enabled, and keep track of larger context. Your job is shifting from writing a perfect prompt to managing a capable junior collaborator.
What AI Is Good At and What It Is Not Good At
Good At
- Turning messy thoughts into a cleaner draft
- Brainstorming options and tradeoffs
- Summarizing, rewriting, translating, and restructuring
- Creating checklists, plans, and step sequences
- Explaining something at different levels
- Working with real inputs like documents, screenshots, pasted notes, and data
- Doing multi-step work when you define the goal, tools, and boundaries
Not Good At
- Reading your mind
- Being perfectly consistent. Ask twice and you may get different answers.
- Being automatically correct. It can sound confident and still be wrong.
- Knowing what changed recently unless it checks current sources
- Understanding your risk tolerance unless you say it clearly
- Keeping secrets safe if you paste them into the wrong tool
- Making high stakes decisions for you
Model Choice Cheat Sheet
Do not use the same model for every job. Match the model to the task, the risk, and the amount of context.
Use for quick rewrites, summaries, small lists, formatting, and low-risk drafts.
Use for planning, analysis, math, debugging, tradeoffs, and answers where being wrong costs something.
Use for long documents, codebases, strategy, complicated synthesis, and important final drafts.
Use when the real input is a screenshot, chart, slide, photo, UI, scan, or document layout.
Use when the job requires actions: browsing, coding, using apps, creating files, or checking results.
Use purpose-built tools when accuracy depends on a database, calendar, spreadsheet, CRM, IDE, or workflow system.
Personal Privacy Traffic Light
These are the rules for personal and consumer AI tools. Do not assume your personal chats are private by default. Settings differ by vendor and plan. If you would be upset to see it leaked, treat it carefully.
- Passwords, private keys, recovery codes, or private links
- Personal identifiers
- Medical or legal details you have not anonymized
- Financial account numbers or full payment details
- Home address, phone number, or full legal name plus birthdate
- Private family, health, or legal notes
- Anything that would hurt if it leaked
- Private photos or sensitive personal messages
- Non-public business or client information
- Drafts you would not want retained or reused
- Public information
- Your own general notes
- Content you would post publicly
- General questions without identifying details
- Anonymized examples you would be willing to share on the internet
Work Privacy Traffic Light
What if we use a paid version at work?
Use these guidelines only if your company plan keeps chats private, has the right admin settings, and says your data is not used to train or improve the model by default for that work plan.
See Vendor Differences with Data Privacy for a plain-English comparison.
- Passwords, API keys, or access codes
- Full card numbers or bank account numbers
- Government ID numbers or full medical records
- Anything you are legally required to keep private
- Internal roadmaps, budgets, or pricing drafts
- Contracts, employee notes, or customer details
- Incident reports without credentials
- Anything that needs legal or policy review before sharing
- Internal how-to docs and process notes
- Code snippets without secrets
- Meeting summaries with identifying details removed
- Approved content such as blog drafts or job listings
- General work questions without personal or confidential details
Vendor and Model Differences
Model names change fast. The stable habit is to choose by capability: speed, reasoning, context, tools, images, privacy controls, and whether the model can use current sources. As of May 2026, the frontier families worth knowing include OpenAI GPT-5.5, Google Gemini 3.5, Anthropic Claude Opus 4.7, and xAI Grok 4.3.
| What to check | Why it matters | How to prompt around it |
|---|---|---|
| Reasoning depth | Hard tasks need planning, evaluation, and correction, not just fluent text. | Ask for assumptions, tradeoffs, a recommendation, and what would change the answer. |
| Context window | Long context helps with documents, codebases, transcripts, and projects with history. | Give the source material and ask the model to ground the answer in it. |
| Tool use | Some models can browse, write files, run code, use apps, or call functions. | Define allowed actions, boundaries, and when it must ask before doing something visible. |
| Multimodal input | Screenshots, charts, scans, photos, and UI bugs need visual understanding. | Ask it to separate observation from interpretation before recommending action. |
| Privacy and training settings | Consumer, team, enterprise, and API products can handle data differently. | Check the plan settings before pasting private, client, medical, legal, or company data. |
Official reference points checked for this local draft: OpenAI latest model guide, Google Gemini 3.5 announcement, Anthropic Claude Opus, and xAI Grok 4 docs.
The Boss Rules
You are the boss of the LLM. The better boss you are, the better responses you will get. These are practical prompt rules for beginners and improving users.
Role: [ROLE]
Goal: [GOAL]
Output format: [FORMAT]
Constraints: [CONSTRAINTS]
Ask me questions if anything is missing.Context:
- Background: [WHAT IS GOING ON]
- Audience: [WHO THIS IS FOR]
- What good looks like: [EXAMPLE OR CRITERIA]
- What to avoid: [LIST]
It is okay if my context is a little messy.
Use it to understand what I really mean.Do not make things up.
Before you answer, ask me the questions you need.
Do this before you draft anything.
Focus on:
- the goal
- the audience
- the constraints
- any missing data
Ask your questions as a numbered list. Then wait.Take your time.
Accuracy matters more than speed.
Work in rounds:
1. Clarify the task.
2. Draft an answer.
3. Critique the draft.
4. Revise the answer.
Surface any assumptions, uncertainty, or missing information.
If anything is unclear, ask me questions first.We are going to do this in rounds.
Round 1: Ask clarifying questions (max 7).
Round 2: Propose 2 to 3 approaches and recommend one.
Round 3: Draft the output.
Round 4: Critique your draft against the goal and constraints.
Round 5: Produce the improved version.Draft the output.
Then critique it against:
- the goal
- the constraints
- what is missing
- what is unclear
Then produce an improved version.Give me options.
Generate 10 ideas.
Pick the best 3.
Explain why they win.
Then ask me which direction to take.You are my writing assistant.
Goal: write in my voice.
Here are 3 to 5 writing samples (paste them below):
[SAMPLE 1]
[SAMPLE 2]
[SAMPLE 3]
First, extract a voice guide with:
- sentence length and rhythm
- typical phrases I use
- what I avoid
- how I open, transition, and close
Do not use emojis, hyphenated sentences, or other clear "AI-written" tells.
I want this to sound like I wrote it and I do not use those things.
Then write [THING] in that voice.
If you are unsure, ask me questions.Before you answer, do this:
1) List any contradictions in my instructions.
2) List what is missing.
3) Rewrite my prompt to be tighter and clearer.
4) Ask me the questions you still need.
Then wait.Prompt Templates
Pick a template. Fill in the brackets. Run it. Then improve it using the boss rules above.
Template 0: The Boss Prompt
Use this when you do not know where to start. It is the safest default template for most tasks.
You are my assistant.
Role: [ROLE]
Goal: [GOAL]
Context:
[PASTE CONTEXT]
Output:
- Format: [FORMAT]
- Length: [LENGTH]
- Tone: [TONE]
Rules:
- If anything important is missing, ask me up to 7 questions before you answer.
- List the assumptions you are making.
- If you are unsure, say so. Do not make things up.
- Use plain English.
- When useful, give me 2 to 3 options and recommend one.
- If this depends on current facts, use current sources, link them, and include the date checked.
- If this would be better with a file, screenshot, or document, ask me for it.
Start by asking your questions. Then wait.Before you answer, do this:
1. List any contradictions in my instructions.
2. List what is missing.
3. Rewrite my prompt to be tighter and clearer.
4. Ask me the questions you still need.
Then wait.Template 1: Persona and Expertise
Use this to set tone, depth, and perspective without pretending the model is magically authoritative.
Respond from the perspective of a [primary role] with expertise in:
- [Specific qualification]
- [Specific experience]
- [Specific specialty]
What this role typically cares about:
- [Priority 1]
- [Priority 2]
Task: [specific task or question]
Instructions:
- Use the perspective, tradeoffs, and priorities of this role.
- Do not invent credentials, facts, or hidden knowledge.
- If experts in this role would disagree, say where they would disagree.
- If anything important is missing, ask me questions first.
Output:
- Terminology level: [professional | accessible]
- Tone: [tone]
- Format: [format]Respond from the perspective of a clear, practical communication coach with expertise in writing firm but respectful emails.
What this role typically cares about:
- clarity
- respect
- getting to the point
Task: Write an email to my landlord about a leaking ceiling that needs repair.
Instructions:
- Use short sentences and plain English.
- Do not use hype.
- If anything important is missing, ask me questions first.
Output:
- Terminology level: accessible
- Tone: confident and respectful
- Format: subject line + email body + 3 alternate subject linesGive me 3 versions.
Explain what each version optimizes for.
Recommend one and tell me why.Template 2: Expert Panel
Use this when you want multiple perspectives, disagreement, and synthesis. Treat it as simulated viewpoints, not a substitute for real experts.
Give me multiple expert perspectives on [problem].
Simulate a panel discussion with the following roles:
- Expert 1: [Specific role] with expertise in [domain]
- Expert 2: [Specific role] with expertise in [domain]
- Expert 3: [Specific role] with expertise in [domain]
Instructions:
- Base each perspective on common real world priorities for that role.
- Do not invent facts.
- If the answer depends on current facts, use sources and link them.
For each expert:
- State their perspective and priorities
- Analyze the problem from that viewpoint
- Propose a solution
Then provide:
- Where they agree
- Where they disagree
- What is uncertain
- A recommended path forward
Format as:
Expert 1:
[response]Work in rounds:
1. Clarify the problem.
2. Draft each viewpoint.
3. Summarize agreement, disagreement, and uncertainty.
4. Revise the recommendation.Template 3: The Fact Checker
Use this when accuracy matters. It forces the model to verify instead of just sounding correct.
You are a tireless researcher and fact checker.
Fact check the text below.
Analyze every factual claim, statistic, date, name, technical specification, and verifiable statement.
Use current sources when needed.
For each important claim, return:
- Claim
- Status: Verified / Unverified / Contradicted
- Source link
- Date checked
- Short note
Then include:
- Claims that still need verification
- Vague or misleading statements
- Internal contradictions
- What is confirmed vs inferred
- Overall risk level: low / medium / high
If you cannot verify something, label it Unverified.
Do not guess.
Text to fact check:
[PASTE YOUR TEXT HERE]Use sources I can click.
Prefer primary sources when possible.
If source quality matters, ask me what sources I trust.Template 4: The Procedural Framework
Use this to get a repeatable process, not just advice.
You are an expert in [domain] and an experienced teacher.
Create a step by step guide to accomplish [specific goal].
Instructions:
- If laws, policies, software, pricing, or product details may have changed, verify current information first.
- Ask questions if anything important is missing.
- Use plain language unless I ask for technical detail.
Structure your response as:
Requirements:
[what is needed]
Preparation:
[setup steps]
Instructions:
[step by step actions]
Troubleshooting:
[common issues and fixes]
Variations:
[alternate approaches]
Goal:
[specific objective]We are going to do this in rounds.
Round 1: Ask clarifying questions.
Round 2: Propose 2 approaches and recommend one.
Round 3: Draft the steps.
Round 4: Critique the steps for gaps and risk.
Round 5: Produce the improved version.Template 5: Navigating a Government System
Use this to turn bureaucratic confusion into practical steps and known pitfalls.
You are an expert at government processes and navigating bureaucracy.
I need help with: [what you are trying to do]
Use current official sources and link them.
Prefer official agency websites.
Tell me which parts vary by jurisdiction.
If anything important is missing, ask me questions first.
My details:
- Who I am: [details]
- Goal: [goal]
- Agency or system: [if known]
- Jurisdiction: [country / state / city / institution]
- Current status: [what I have done]
- Constraints: [deadlines, missing documents, budget, legal issues]
- Dates and deadlines: [known dates]
- Special factors: [anything unusual]
What I want:
1. A simple overview of how the system works
2. A numbered action plan for my situation
3. Required forms, documents, IDs, and evidence
4. Expected fees
5. Decision points and bottlenecks
6. Common mistakes and how to avoid them
7. A realistic timeline
8. Questions the agency is likely to ask me
9. A script for calling or emailing the agency
10. A backup plan if the main path stallsTemplate 6: Test and Interview Prep
Use this when you want practice, feedback, and increasing difficulty.
You are an experienced teacher, coach, and interviewer in [field].
Help me prepare for [role / exam / conversation / topic].
Do the following:
1. Ask me for the relevant materials first.
2. Review the materials and identify focus areas.
3. Create a structured practice session with questions of varying difficulty.
4. After each answer, score me briefly on:
- clarity
- completeness
- credibility
5. Tell me whether I should expand, clarify, or move on.
6. End with a summary of strengths and the top improvements I should make.
Start by asking for my materials.Work in rounds:
1. Ask for the materials.
2. Build the practice plan.
3. Critique the plan for missing difficulty levels.
4. Improve it.Template 7: Specific Templatized Output
Use this when you need output in a strict format every time.
You are excellent at following directions.
I am giving you a strict output template with [PLACEHOLDERS].
Rules:
- Output must match the template exactly.
- Do not add extra sections.
- Do not invent facts.
- If a field is missing, either ask me for it or mark it as [MISSING].
Template:
[PASTE TEMPLATE HERE]
Apply this template to:
[YOUR REQUEST]If the template and my instructions conflict, stop and ask me which one should win.Template 8: Research with Sources
Use this when you need current, source-backed answers instead of generic explanations.
Research this topic using current sources:
[TOPIC]
Instructions:
- Use current sources.
- Link each major claim.
- Include the date checked.
- Separate what is confirmed from what is inferred.
- If sources conflict, say so.
- If you cannot verify something, label it Unverified.
Output:
1. Short answer
2. Key findings with source links
3. What is confirmed
4. What is inferred
5. Open questions or uncertaintyTemplate 9: Document or PDF Review
Use this when you are uploading a file and want the model to stay grounded in that file.
I am uploading a document or PDF.
First:
1. Summarize what the document is.
2. List any ambiguities, missing pages, or unreadable parts.
Then do this task:
[TASK]
Rules:
- Answer from the document first.
- If you use outside knowledge, label it clearly.
- Quote or cite the part of the document you are relying on when possible.
- If the file is not enough to answer fully, tell me what is missing.Template 10: Screenshot or Image Analysis
Use this for screenshots, photos, dashboards, UI bugs, and visual analysis.
I am uploading an image or screenshot.
First:
1. Describe exactly what you see.
2. List any ambiguities or things you cannot read clearly.
Then do this task:
[TASK]
Rules:
- Separate observation from interpretation.
- If text is visible, extract the text first.
- If you are inferring something, label it as inference.
- If the image alone is not enough, tell me what else you need.Template 11: Compare Sources and Resolve Conflicts
Use this when different articles, documents, or vendors disagree.
Compare these sources on:
[TOPIC]
Sources:
- [SOURCE 1]
- [SOURCE 2]
- [SOURCE 3]
Output:
1. What all sources agree on
2. Where they conflict
3. Which source appears strongest and why
4. What is still uncertain
5. Your best current conclusion
Rules:
- Link the sources.
- Prefer primary or official sources when possible.
- Separate evidence from interpretation.Template 12: Structured JSON or Table
Use this when you want data you can paste into a spreadsheet, script, or workflow.
Return the answer in [JSON | table] format.
Schema or columns:
- [FIELD 1]
- [FIELD 2]
- [FIELD 3]
Rules:
- Use null for unknown values.
- Do not invent missing values.
- If the schema is unclear, ask me questions first.
- Keep all field names exact.
Task:
[YOUR REQUEST]Return the answer in JSON format.
Schema:
- company_name
- product_name
- category
- target_customer
- pricing_notes
- source_url
- confidence
Rules:
- Use null for unknown values.
- Do not invent missing values.
- Keep all field names exact.
Task:
Research the top embedded DevOps tools vendors and return the results in this schema.Template 13: Agentic Task
Use this when the model can take actions, use tools, inspect files, browse, or work in an app.
Goal:
[What I want completed]
Allowed actions:
- [Browse / inspect files / write code / create a document / use a tool / other]
Boundaries:
- Do not publish, send, delete, purchase, or make external changes without asking me first.
- Do not expose private information.
- Stay within these files, apps, or systems: [BOUNDARIES]
Inputs:
[Paste notes, links, files, screenshots, examples, or constraints]
Work style:
- Start by checking the relevant context.
- Make a short plan.
- Do the work.
- Verify the result.
- Tell me what changed, what you checked, and what still needs human review.Work in Rounds
Most people stop after one prompt. That is why they get mediocre output. Better models reward better management.
Round 1: Understand
Restate the goal, assumptions, missing context, and likely risks.
Round 2: Plan
Give me a concise plan before doing the work.
Round 3: First pass
Create the first version. Optimize for usefulness, not perfection.
Round 4: Check
Critique your own output. What is missing, weak, unsupported, or risky?
Round 5: Revise
Fix the important issues and give me a cleaner version.
Round 6: Human review
List the claims, decisions, or actions I should verify myself before relying on this.