It learned patterns
AI models are trained on large examples of language, code, images, and other data patterns.
One connected lesson
What we’re going todive into.
Eight focused sections, one smooth path. You'll start with a plain-English foundation of what AI is, then the AI layers, real prompting examples, projects and guardrails, and finish by setting up Claude and ChatGPT to match your own tone and work style. A plain-English glossary sits at the end for reference.
01What is AI?Plain-English foundation: patterns, instructions, context, outputs, and judgment.
02AI EcosystemTruck analogy, model versions, future capabilities, and tools.
03PromptingWeak ask vs strong ask with typing demos.
✦AI ImagesPrompt for images; gpt-image-1 vs gpt-image-2.
04Projects & ContextClean workspaces, files, examples, and token awareness.
05GuardrailsRead, draft, send, delete, execute with controls.
062025 → 2027How companies mature from ad hoc to governed.
07Claude setupPreferences, styles, projects, and instructions.
08ChatGPT setupCustom Instructions, Memory, and default tone.
09Key termsPlain-English glossary: hallucination, tokens, context, guardrails, agents.
Section 1
What is AI? Start with the plain-English version.
Before model names, tools, and settings, here is one shared definition: AI is software that uses patterns, instructions, and context to generate useful outputs or help complete tasks. It is powerful, but it still needs human direction and review.
AIA smart assistant that follows your instructions.It drafts, summarizes, compares, answers questions, and helps you get work done.
You give instructions
Your prompt tells the assistant what job to do, what tone to use, and what result you want.
Context shapes the answer
Files, chat history, examples, and details make the output more specific and useful.
Humans judge the result
AI can be confident and still be wrong. People still own accuracy, safety, and decisions.
AI is not magic
It does not “know” like a person. It generates likely, useful responses from patterns, instructions, and available context.
AI is not one product
ChatGPT, Claude, Gemini, Copilot, and Perplexity are different assistants. They may use different models, tools, memory, files, and web access.
AI is a work partner
The best mental model is a capable assistant: fast, helpful, and flexible, but better with clear instructions and good supervision.
Good atSummariesDraftsBrainstormingComparisonsAnalysisImagesCoding help
Needs care withFactsLegal or medical advicePrivate dataPermissionsFinal decisionsOutdated information
What AI is for
AI gets you to a better first draft, a clearer understanding, or a better next step faster. It does not remove your responsibility. It helps you think and produce.
Don't over-trust it
AI is not always right, not private by default, and not ready to take action without controls. Aim for useful confidence, not blind trust.
Section 2
AI Ecosystem: explain it with a real truck.
The easiest way to stop confusion is to separate the layers. Company, assistant, model, mode, prompt, project, files, and connectors are not the same thing.
Start with the question
“Who built the truck, what do you drive, what powers it, and what is attached to it?”
Map the layers
OpenAI, Anthropic, Google, Microsoft, and Meta are companies. ChatGPT, Claude, Gemini, Copilot, and Perplexity are assistants. Models power the assistant.
Same truck, different job
Different engine, drive setting, cargo, and attached tools change the result. That is how AI systems behave differently.
Live example
Same question. Better setup.
A vague question about AI tools produces generic advice. A specific job and real context produce a useful recommendation.
Weak ask
Typing prompt
What is the best AI?
Typical answer: “It depends. ChatGPT, Claude, Gemini, and Perplexity are all good options.”
Better ask
Typing prompt
I run a small business. I need one assistant for customer emails, one for current web research with citations, and one for long policy documents. Recommend a simple starting stack and explain why.
Better answer: “Use a general assistant for daily drafting, a research-first assistant when citations matter, and a long-document assistant for contracts, policies, and project files.”
Model versions keep moving
Do not memorize model names. Learn what usually changes.
When a model moves from one version to the next, the name is only the label. The real question is: what changed in capability, cost, speed, context, tools, and safety?
Better work quality
Newer versions usually improve reasoning, writing, coding, instruction following, analysis, and reliability on harder tasks.
More ways to work
Expect stronger multimodal work: text, images, audio, video, files, screens, spreadsheets, and connected tools inside one workflow.
More context and control
Newer versions can hold more in mind at once, handle longer documents, and use more tools.
No fixed release calendar
Model upgrades do not arrive on a tidy monthly schedule. Teach people to check release notes and expect names to change.
Older versionGood for everyday work.May need more prompting, more checking, and more context cleanup.
→
Newer versionUsually better at harder work.May cost a bit more or have different limits.
Models keep getting more capable, more connected, more multimodal, and more action-oriented. That makes prompting, context, permissions, and judgment more important, not less.
Section 3
Prompting: better inputs create better outputs.
You feel this one immediately. Weak prompts make AI guess. Strong prompts give the assistant a role, context, facts, tone, constraints, and an output format.
Typing demo
Weak prompt vs strong prompt
Weak prompt
You type
Write an email to the customer about the delay.
AI returns
Subject: Update on delay
Hi, there is a delay. We are working on it and will let you know when we have more information. Thanks.
Hi, there is a delay. We are working on it and will let you know when we have more information. Thanks.
- No customer name
- No timeline
- No reassurance
- No next step
Strong prompt
You type
Write a short, calm email to Sarah explaining cabinet delivery will delay the final walkthrough by four business days. Confirm painting, flooring, fixtures, and appliances are complete. Give the new walkthrough date: Wednesday, May 28 at 10:00 AM. Avoid blaming the vendor. Invite her to call or text me with questions.
AI returns
Subject: Final walkthrough rescheduled
Hi Sarah, I wanted to let you know cabinet delivery will push the final walkthrough by four business days. The good news is the painting, flooring, fixtures, and appliances are complete. Your new walkthrough is Wednesday, May 28 at 10:00 AM. Please call or text me with any questions.
Hi Sarah, I wanted to let you know cabinet delivery will push the final walkthrough by four business days. The good news is the painting, flooring, fixtures, and appliances are complete. Your new walkthrough is Wednesday, May 28 at 10:00 AM. Please call or text me with any questions.
- Clear facts
- Right tone
- Specific date
- Useful next step
Task
What should AI do?
Context
What background matters?
Audience
Who will read it?
Tone
How should it sound?
Rules
What must it include or avoid?
Format
What should it look like?
The mental model
Think of them like shape-shifting twins.
On their own, ChatGPT and Claude are generic. Name the shape you need in your prompt, the role, the audience, the tone, the format, and they become exactly that: a patient teacher, a blunt editor, a careful analyst. Your instructions are what give them form. Vague prompt, vague shape. Specific prompt, specific shape.
AI Images
AI can make images too.
Tools like ChatGPT can create images from a description. It works just like prompting text: the more specific you are about the subject, the style, and the details, the closer you get to what you pictured.
Subject
What is in the picture. "A friendly golden retriever on a front porch."
Style
How it should look: photo, watercolor, cartoon, line drawing, product shot.
Details
Colors, lighting, mood, setting, and anything that must be there.
Format
Shape and use: square for social, wide for a banner, room for text.
Weak vs strong image prompt
Say what you actually want.
A vague image prompt gives you a generic, stock-looking picture. A specific one gives you something you can use.
Weak prompt
Make a picture of a coffee shop.
Likely result: A generic, could-be-anywhere coffee shop.
Strong prompt
A warm neighborhood coffee shop at golden hour, photo style, wooden tables, a chalkboard menu, soft window light, with space at the top for a headline.
Likely result: A usable, on-brand image you could put on a flyer.
Two image models
gpt-image-1 vs gpt-image-2.
Just like text models, image models come in versions. Newer usually means sharper detail, better text inside the image, and closer prompt-following, sometimes for a bit more cost or time.
gpt-image-1
The earlier version. Fast and solid for simple images and quick drafts.
gpt-image-2
Newer. Sharper detail, better at text in the image, and follows the prompt more closely.
Section 4
Projects, Files & Context: clean workspace, better answers.
AI does better when the work is organized. A focused project gives it the right instructions, trusted files, examples, and history.
Context order
The assistant builds on what it can see.
1System instructions
2Your prompt
3Prior conversation
4Uploaded files
5AI response
Tokens and limits
Tokens are the AI’s working memory meter.
Tokens are chunks of text. Your prompt, the assistant’s reply, conversation history, uploaded-file text, and tool results all take space. More context can be helpful, but long messy chats consume more capacity and can make the work slower or less focused.
Clean promptlow token load
Long chat + fileshigher token load
Can you change models if you exhaust tokens?
Sometimes, yes. If you hit a limit on a premium model, the app may let you switch to a lighter one. If you hit a bigger account or plan limit, switching may not help. The fix is usually simple: start a fresh chat, use a project, or wait for the limit to reset.
Plain-English rule of thumb: 1 token is roughly 4 English characters, and 100 tokens is roughly 75 English words.
Real example
Same request. Different workspace.
Uploading a file once somewhere is not the same as giving the assistant clean context inside the right project.
Messy chat
“Summarize this for the board.”
The assistant does not know which file is current, what the board cares about, or what decision is being made.
Clean project
“Use the attached June finance report and board template.”
The assistant summarizes variance, risks, decisions needed, and next steps in the format leadership expects.
What belongs
Purpose, rules, reference files, examples, active chats, and shared templates.
What hurts
Unrelated topics, duplicate files, old versions, unclear instructions, and confidential data that does not belong.
Best practice
One project per workstream, clean file names, current sources, handoff summaries, and fresh chats when messy.
Section 5
Actions, Permissions & Guardrails: power needs protection.
When AI can only read or draft, the risk is lower. When it can send, delete, buy, change records, or run commands, you need approval, logging, rollback, and clear ownership.
Real failure scenario
Bad automation can move fast in the wrong direction.
AI taking action is not the problem. Unreviewed, over-permissioned action is.
Risky instruction
Archive all old project files and clean up the customer folder.
Risk: The assistant may move active files if “old” and “project” are not clearly defined.
Safer instruction
List files that appear inactive for 24+ months. Do not move or delete anything. Create a review table with file name, last modified date, owner, reason, and recommended action. Wait for approval.
Safer result: AI prepares the review. A human approves before anything changes.
Least privilege
Give the assistant only the access it needs for the job.
Human approval
Require sign-off before sending, deleting, purchasing, or changing records.
Logging
Know who did what, when, and why.
Rollback
Keep versions and recovery paths so mistakes can be fixed.
Section 6
2025 → 2027: how businesses should mature.
For business owners and managers, the goal is to move from ad hoc experimentation to organized, connected, governed AI that produces measurable outcomes.
Maturity path
Business example
Policy turns random AI use into repeatable work.
There is a real difference between people experimenting and the company operating with AI.
Ad hoc
“Everyone use whatever AI helps.”
Personal accounts, copied client data, no review, no shared prompts, no way to measure whether it helped.
Organized
“Use approved tools and shared project spaces.”
Clear acceptable use, reusable prompts, no sensitive data in random tools, manager review, and simple outcome tracking.
Section 7
Claude setup: make the assistant sound like you.
Set the defaults before you ask for work. Two things shape Claude's voice: your profile preferences plus a Style for how it writes, and Projects with project instructions for work-specific context.
Setup checklist
- Open Settings, then Profile. Add what Claude should call you and your personal preferences for how it responds.
- Pick a Style in the chat (Normal, Concise, Explanatory, or Formal), or create a custom Style from your own writing.
- Create a Project for each workstream and add project instructions plus any reference files.
- Test the voice with a short email, a summary, and a plan before you use it for real work.
Claude Projects are designed as self-contained workspaces with chat history, knowledge, and project instructions. Usage is affected by conversation length, files, model choice, and features, so clean projects help conserve capacity.
Profile prompt example
Give Claude a default way to respond.
Copy a version of this into your Claude profile preferences (or a project's instructions), then adjust it to fit your work.
No setup
Write this in a better way.
Likely result: Polished, but generic. It may sound like AI because it has no voice target.
Claude preference
I run a small business. When you respond to me, use plain English: warm, direct, and practical. Short paragraphs, no corporate buzzwords, and no em dashes. For emails, sound like a helpful business owner, not a marketing department. Give me the usable version first.
Likely result: Shorter, warmer, more personal, and easier to use immediately.
Menu labels may change over time, but the pattern holds: preferences and a Style for voice, Projects for work context.Section 8
ChatGPT setup: Custom Instructions, Memory, and tone.
Set the defaults once so every new chat starts closer to what you need. Personalization is not magic. It is standing context plus memory controls.
Setup checklist
- Open Settings from your profile menu.
- Select Personalization and enable customization.
- Add Custom Instructions for your role, tone, formatting rules, and things to avoid.
- Review Memory so ChatGPT remembers useful context without keeping stale or sensitive details.
- Test it with a real email, summary, and decision memo.
OpenAI says Custom Instructions are available on Web, Desktop, iOS, and Android and are applied immediately across chats. ChatGPT plans may also include model/message limits, so use heavier models when the work truly needs them.
Official reference image
Personalization is a real setting.
This is the Personalization and Memory settings area in ChatGPT, so you know exactly where to look.
Custom Instructions example
Make your default voice explicit.
A one-time instruction changes every future draft.
Blank settings
Make this sound better.
Likely result: Safe, generic, and often too formal.
Custom Instructions
I run a small business. Write in my voice: short, warm, direct, plain English. Avoid hype, em dashes, corporate buzzwords, and long bullet lists. Give me the usable version first, then explain only if needed.
Likely result: More consistent tone, less rework, and faster drafts.
LearnAI
Wrap-up
Practical. Relevant. Actionable.
By the end, you can explain what AI is in plain English, choose the right assistant, write stronger prompts, structure projects better, explain AI risk clearly, mature your business's AI use, and set up Claude and ChatGPT for your own tone.
Reference
Key terms, in plain English.
A quick glossary of the words that come up around AI. Skim it now, or come back when a term shows up.
Hallucination
When AI states something false with full confidence. It is filling a gap with a likely-sounding answer, not lying.
Watch for invented citations, cases, names, or numbers. Always verify facts and quotes.
Token
The small chunks of text AI reads and writes. Your prompt, your files, and the reply all spend tokens.
Rough rule: 1 token is about 4 characters, and 100 tokens is about 75 words.
Context window
How much the model can hold in mind at once: your prompt, your files, and the conversation so far.
A long, messy chat pushes older details out of view. Start fresh when it drifts.
Prompt vs system prompt
Your prompt is what you type now. The system prompt, or custom instructions, is the standing setup that shapes every reply.
Set "plain English, no em dashes" once and it applies to every chat.
Temperature
A creativity dial. Low is safe and predictable. High is more varied and surprising.
Keep it low for policy and numbers, higher for brainstorming.
Guardrails
The limits that keep AI safe and on task: refusals, permissions, approvals, and your own rules.
"List files for review, do not delete anything, wait for approval." See the Guardrails section.
Grounding (RAG)
Giving AI real source material to answer from instead of relying on memory.
Attach the actual policy so it quotes your document, not a guess.
Agent
AI that can take steps and use tools, not just chat. It can search, open files, and draft a reply.
This is why permissions and review matter even more.