AI Skills Everyone Should Learn

8 evergreen AI literacy skills — practical, not technical. These are the skills that matter regardless of which AI tool you use or how the technology evolves.

The 8 skills

How to use these skills

• Read one skill per day (8 days total)
• Practice each skill with a real task
• Bookmark these pages — you'll return to them

Why these skills matter

AI tools will change. ChatGPT may be replaced. New tools will appear. But these 8 skills are tool-agnostic — they'll serve you regardless of what AI looks like in 5 years.

Last reviewed: 2026-08

Asking Better Questions

The single most important AI skill. The quality of AI's answer is determined by the quality of your question.

Why this skill matters

Most people are disappointed with AI because they ask bad questions. They type "write me a marketing email" and get generic garbage back. Then they blame AI.

But AI is a mirror — it reflects the quality of your input. Ask a vague question, get a vague answer. Ask a specific question with context, format, and examples, and you'll get something useful.

This skill is tool-agnostic. Whether you use ChatGPT, Claude, Gemini, or some tool that doesn't exist yet, asking better questions will always help.

The anatomy of a bad question

Bad questions share these traits:

• Too vague: "Help me with my resume"
• Too broad: "Tell me about AI"
• No context: "Write an email"
• No format: "Give me ideas"
• No audience: "Make it good"

The anatomy of a good question

Good questions include:

1. Context — who you are, what you're trying to do

Bad: "Write an email"
Good: "I'm a freelance graphic designer following up with a potential client I met at a conference last week. Write a follow-up email."

2. Specific task — exactly what you want

Bad: "Help me with my resume"
Good: "Rewrite the bullet points in my Experience section to be more impact-focused, using the STAR framework (Situation, Task, Action, Result)."

3. Format — how you want the output

Bad: "Give me ideas"
Good: "Give me 10 blog post ideas about productivity for ADHD adults, formatted as a numbered list with a one-sentence premise for each."

4. Constraints — length, tone, things to avoid

Bad: "Write a summary"
Good: "Summarize this 50-page report in under 300 words, in plain language a non-technical executive could understand. Avoid jargon."

5. Examples — show what good looks like

Bad: "Write a tweet"
Good: "Write a tweet announcing my new course. Here's an example of a tweet I liked that has a similar tone: [paste example]."

The 5W1H framework

Before asking AI anything, ask yourself:

• Who am I asking this for?
• What specifically do I want?
• When is the context (timing, deadline)?
• Where will this be used (email, blog, presentation)?
• Why am I asking (what's the underlying goal)?
• How do I want the answer formatted?

The "one more sentence" trick

When you've written your question, ask: "Could I add one more sentence of context that would make AI's answer 10% better?" Usually, the answer is yes. Add it.

Example progression:

1. "Write me a blog post about AI"
2. "Write me a 500-word blog post about AI for small business owners"
3. "Write me a 500-word blog post about how small business owners can use AI to save time, with 3 specific examples, conversational tone, for an audience that's curious but skeptical"

Each iteration is 10x more useful.

Common question-killing phrases to avoid

• "Just..." (signals you don't really care)
• "Quick question..." (signals low effort)
• "Whatever you think..." (abdicates your judgment)
• "Make it good..." (vague to the point of useless)
• "Something about..." (you don't know what you want)

Practice exercise

Take this bad question and improve it using the 5 elements above:

"Help me with my presentation"

Improved version:

"I'm presenting to 30 high school teachers about using AI in their classrooms. The presentation is 20 minutes long, including 5 minutes for Q&A. I want to cover 3 practical AI tools teachers can use tomorrow, with one concrete classroom example for each. The tone should be encouraging, not technical — these teachers are AI-skeptics. Format the output as a slide-by-slide outline with timing for each slide."

The meta-skill: asking AI to help you ask better

You can use AI to improve your questions. Try this:

"Here's my question to you: [your draft question]. Before answering, suggest 3 ways I could improve this question to get a better answer from you."

AI will tell you what context it wishes you had provided.

Recap

• Bad questions get bad answers. AI is a mirror.
• Good questions include: context, specific task, format, constraints, examples.
• Use the 5W1H framework before asking.
• Add "one more sentence" of context — it's almost always worth it.
• Avoid question-killing phrases like "just" and "whatever you think."
• Use AI itself to improve your questions.

Memory trick

CCFCE — Context, Constraints, Format, Examples, (specific) Ask.

Or simpler: "Who, What, When, Where, Why, How" — the journalist's questions still work.

Last reviewed: 2026-08 · Skill 1 of 8

Choosing the Right AI Tool

ChatGPT, Claude, Gemini, Perplexity, NotebookLM, Midjourney, Otter, Gamma... With so many tools, how do you choose? This skill gives you a decision framework.

Why this skill matters

There's no single "best" AI tool. Different tools are good at different things. Using the wrong tool wastes time and produces worse results.

The good news: you don't need to memorize every tool. You just need a decision framework.

The decision framework

Ask yourself these questions in order:

Question 1: What type of task?

See Tools Catalog for specific tool recommendations in each category.

Question 2: What are my constraints?

• Budget: Free vs. paid
• Privacy: Can I use a cloud tool, or do I need local/self-hosted?
• Integration: Does it need to work with my existing apps?
• Language: Does it support Bengali / my language?
• Device: Mobile vs. desktop
• Offline: Do I need it to work without internet?

Question 3: What's the stakes?

• Low stakes (drafting, brainstorming): Any tool works
• Medium stakes (work emails, content): Use a reliable tool, verify output
• High stakes (medical, legal, financial): Use specialized tools, verify everything, consult experts

Quick decision guide

"I want to chat with AI about anything"

→ Start with ChatGPT (most versatile) or Gemini (best free tier, Google integration)

"I want to summarize a long document"

→ Claude (200K+ context) or NotebookLM (for your own PDFs)

"I want to research with citations"

→ Perplexity (every answer cited to a source)

"I want current information"

→ Perplexity or Gemini (real-time web access)

"I want to write better"

→ ChatGPT/Claude for drafting + Grammarly for editing

"I want to make an image"

→ Bing Image Creator (free) or Midjourney (premium quality)

"I want to make a presentation"

→ Gamma (AI-generated) or Canva (template + AI)

"I want to transcribe a meeting"

→ Otter.ai (most popular) or Whisper (free, open-source)

"I want to translate"

→ DeepL (best quality) or Google Translate (most languages)

"I want to write code"

→ GitHub Copilot (in IDE) or Cursor (AI-first editor) or ChatGPT/Claude (one-off help)

"I want to learn something"

→ ChatGPT (custom GPT tutor) or Khanmigo (K-12) or Duolingo Max (languages)

The 2-tool minimum

For most people, 2 tools cover 90% of needs:

1. A general chatbot (ChatGPT, Claude, or Gemini) — for everything
2. Perplexity — for anything needing current info or citations

That's it. Don't over-tool.

The 5-tool power user

If you use AI heavily, add:

1. NotebookLM — for working with your own documents
2. An image generator — Bing (free) or Midjourney (paid)
3. Otter.ai or equivalent — for meeting transcription

More than 5 tools = you're probably overcomplicating.

Tool comparison: ChatGPT vs Claude vs Gemini

When to switch tools

Switch when:

• You hit usage limits on your current tool
• You need a capability your tool lacks (e.g., Claude for long docs)
• Quality drops on specific task types
• You need real-time info (switch to Perplexity/Gemini)

Don't switch just because:
- A new tool launched (most are hype)
- Your friend uses something different
- You're bored

Trying a new tool

When evaluating a new AI tool:

1. Start with free tier — never pay before testing
2. Test on a real task — not "hello world"
3. Compare to your current tool on the same task
4. Check privacy policy — what do they do with your data?
5. Read independent reviews — not just the company's marketing
6. Give it 2 weeks — first impressions are misleading

The "good enough" principle

Most AI tools are "good enough" for most tasks. The differences between ChatGPT, Claude, and Gemini are smaller than marketing suggests.

Don't optimize for the "best" tool. Pick one, learn it deeply, and only switch when you hit a real limitation.

Recap

• Different tools are good at different things
• Use the decision framework: task type → constraints → stakes
• 2 tools cover 90% of needs: a chatbot + Perplexity
• 5 tools for power users: chatbot + Perplexity + NotebookLM + image gen + Otter
• Don't over-tool — pick one, learn it deeply
• When evaluating new tools: free tier first, real task test, compare to current

Memory trick

"Chat-Perp-Note-Image-Meet" — Chatbot, Perplexity, NotebookLM, Image generator, Meeting tool.

That's the 5-tool power user stack.

Last reviewed: 2026-08 · Skill 7 of 8

Collaborating with AI, Not Depending on It

AI is a powerful collaborator but a dangerous crutch. This skill teaches you how to use AI without losing your own abilities.

Why this skill matters

There's a quiet danger in AI use: skill atrophy. If you use AI to write every email, your writing skills weaken. If you use AI to do every calculation, your numeracy fades. If you use AI to make every decision, your judgment dulls.

This isn't hypothetical. Studies are already showing measurable skill declines in heavy AI users.

The goal isn't to avoid AI. It's to use AI as a collaborator — one that amplifies your abilities rather than replacing them.

The collaboration vs. dependence test

How do you know if you're collaborating or depending? Ask:

Collaborating:
- AI helps me do things faster, but I could do them myself
- I learn from AI's suggestions and improve over time
- I verify AI's output and make my own judgment
- I use AI for tasks I understand, not tasks I don't
- If AI disappeared tomorrow, I'd be slower but functional

Depending:
- I use AI for tasks I don't understand and don't try to learn
- I trust AI's output without verifying
- I can't do the task without AI
- I copy-paste AI output without reading it
- If AI disappeared tomorrow, I'd be stuck

The 3 levels of AI use

Level 1: AI as tool (collaborating)

You use AI for specific tasks, verify the output, and learn from the interaction. Your skills grow over time.

Example: You write a draft email, ask AI for improvement suggestions, consider the suggestions, and apply the ones that make sense. Over time, your writing improves because you're learning from AI's feedback.

Level 2: AI as assistant (slight dependence)

You delegate substantial work to AI but still review and own the output. Your skills plateau — you're not improving, but not declining.

Example: You ask AI to draft the email from scratch, then edit it. The email gets sent, but you're not actively learning to write better emails. Over time, your writing skill stays the same.

Level 3: AI as crutch (heavy dependence)

You rely on AI for tasks you don't understand. You can't do the task without AI. Your skills actively decline.

Example: You ask AI to draft the email, you don't really read it, you hit send. If asked to write the email yourself, you couldn't. Over time, your writing skill atrophies.

Goal: Stay at Level 1, occasionally use Level 2, never reach Level 3.

The atrophy audit

For each AI use, ask: "Am I using this to learn, or to avoid learning?"

Skills worth protecting

Some skills are worth maintaining even when AI can do them:

1. Writing

Writing is thinking. If you can't write, you can't think clearly. Use AI to improve your writing, not replace it.

Practice: Write first drafts yourself. Use AI for editing and feedback.

2. Critical thinking

The ability to evaluate claims, spot flaws, and form judgments. AI can't do this for you — it can only suggest.

Practice: Question AI's output. Look for flaws. Form your own opinion before asking AI.

3. Numeracy

Basic math and statistical reasoning. If you can't sanity-check AI's numbers, you'll be misled.

Practice: Do quick calculations yourself. Use AI for complex ones, but verify the result.

4. Creativity

Generating original ideas. AI is good at combining existing ideas, but genuine originality still comes from humans.

Practice: Brainstorm without AI first. Then use AI to expand and refine.

5. Empathy and judgment

Understanding people, reading situations, making human judgments. AI can't do this.

Practice: Make your own decisions about people. Don't use AI to analyze relationships or make personal choices.

The "do it yourself first" rule

For any task you want to stay skilled at, do it yourself first. Then use AI to improve.

Workflow:
1. Attempt the task yourself (write the draft, solve the problem, make the decision)
2. Use AI to critique, suggest, or improve
3. Compare your version to AI's
4. Learn from the differences
5. Apply the improvements that make sense

This workflow keeps you actively engaged, so your skills grow rather than atrophy.

When it's OK to depend on AI

You can't be an expert at everything. It's fine to depend on AI for:

• Tasks outside your core skills (e.g., a writer using AI for math)
• Tasks you'll never need to do without AI (e.g., grammar checking)
• Low-stakes tasks where learning isn't worth the time
• Tasks where AI is genuinely better (e.g., summarizing long documents)

The danger is when you depend on AI for core skills — the things that define your work or life.

The skill maintenance plan

For each core skill you want to maintain:

1. Identify the skill (writing, coding, analysis, etc.)
2. Decide frequency (daily, weekly, monthly)
3. Practice without AI for that frequency
4. Use AI to learn from the practice
5. Track improvement over time

Example: "I will write one blog post per week without AI, then use AI to critique it and learn."

Signs you're depending too much

• You feel anxious when you can't use AI
• You can't explain your own work
• You're surprised by AI's output (you didn't know what it would say)
• Your work without AI is noticeably worse than with AI
• You've stopped learning new things in your field
• You defer to AI even when you have expertise

What to do if you're depending too much

1. Acknowledge it — denial is the first obstacle
2. Identify which skills are atrophying
3. Gradually reduce AI use for those skills
4. Practice without AI regularly
5. Use AI to learn, not to do
6. Be patient — skill rebuilding takes time

The long view

AI is getting more capable every year. The skills that AI can replace will expand. But the skills that make us human — judgment, empathy, creativity, wisdom — will become more valuable, not less.

Invest in those skills. Use AI to amplify them, not replace them.

Recap

• Use AI as a collaborator, not a crutch
• Test yourself: "Could I do this without AI?"
• Stay at Level 1 (AI as tool), avoid Level 3 (AI as crutch)
• Protect core skills: writing, critical thinking, numeracy, creativity, judgment
• Use the "do it yourself first" rule for skills you want to maintain
• It's OK to depend on AI for non-core tasks
• Watch for warning signs: anxiety without AI, can't explain your work
• If depending too much: gradually reduce, practice without AI

Memory trick

"Do first, AI second."

For any skill you want to keep, do it yourself first. Then let AI help.

Last reviewed: 2026-08 · Skill 8 of 8

Fact-Checking AI Outputs

AI confidently states false information (hallucinations). This skill teaches you how to verify AI claims without spending hours.

Why this skill matters

In 2023, a lawyer used ChatGPT to write a legal brief. ChatGPT cited several cases — with quotes and dates. The lawyer submitted the brief. The judge discovered: none of those cases existed. The lawyer was sanctioned.

This happens every day, to ordinary people, in less dramatic ways. You ask AI for a recipe, it works. You ask AI for a medical explanation, it sounds right but is wrong. You ask AI for a statistic, it invents a plausible-sounding number.

You must verify anything that matters.

What to verify (and what's safe to skip)

Always verify

• Specific facts: names, dates, statistics, quotes
• Citations: papers, books, court cases, laws
• Numbers: prices, percentages, quantities
• Medical/legal/financial claims: high stakes
• Recent events: anything after AI's training cutoff
• URLs: AI often invents non-existent links
• Code: AI code can have subtle bugs

Usually safe to trust

• General explanations of well-known concepts
• Writing/style suggestions (grammar, tone, structure)
• Brainstorming (idea generation)
• Summaries of text you provide
• Translations (cross-check important ones)
• Coding patterns (but verify the actual code)

The 3-source rule

For any claim that matters, verify against at least one independent source that AI didn't generate. Ideally:

1. Primary source (the original paper, law, document)
2. Reputable secondary source (major news, textbook, encyclopedia)
3. Cross-reference (a different source confirming the same fact)

Fact-checking workflow

Step 1: Identify claims

After AI gives you an answer, list the specific factual claims. Example:

AI said: "The Eiffel Tower was built in 1889 for the World's Fair, designed by Gustave Eiffel, and is 330 meters tall."

Claims to verify:
- Built in 1889
- For the World's Fair
- Designed by Gustave Eiffel
- 330 meters tall

Step 2: Use a search engine

Google or Perplexity each claim. Don't ask the same AI to verify itself — it may double down on the error.

Step 3: Use Perplexity for cited research

Perplexity is specifically built to provide cited answers. Use it for fact-checking:

"Is it true that [claim]? Provide sources."

Step 4: Use NotebookLM for your own documents

If you're asking about documents you have (a textbook, a contract, a research paper), upload them to NotebookLM. It only answers from your documents, eliminating hallucinations from training data.

Step 5: Verify URLs separately

AI frequently invents URLs that look real but don't exist. Always click links before using them.

Spotting hallucinations

Hallucinations have patterns. Watch for:

1. Plausible-sounding but unverifiable claims

If you can't find the claim anywhere else, it's probably invented.

2. Specific numbers without sources

"73% of users prefer..." — where did that number come from?

3. Citations to non-existent papers

AI often generates realistic-looking academic citations. Search the title in Google Scholar — if no results, it's hallucinated.

4. Confidence on obscure topics

AI is most reliable on well-documented topics. If you ask about something obscure, it may confidently invent.

5. "As of my last training..." deflections

Sometimes AI admits its training cutoff. Sometimes it doesn't. Don't assume.

The "trust, but verify" hierarchy

Trust AI more for:
- Topics well-covered in training data
- Tasks where you can quickly check the output
- Low-stakes decisions

Trust AI less for:
- Recent events
- Obscure topics
- High-stakes decisions (medical, legal, financial)
- Topics with strong misinformation risk (politics, health)

Quick verification tools

When you can't verify

If you can't verify a claim and it matters:

1. Don't use it. Better to omit than to spread misinformation.
2. Hedge explicitly. "AI suggested this, but I couldn't verify it."
3. Ask a human expert. For medical, legal, financial — always.

The cost of NOT verifying

• Academic: failed assignments, plagiarism accusations, damaged reputation
• Professional: embarrassing emails, lost deals, fired for fake citations
• Personal: bad decisions based on wrong info, health risks from wrong advice
• Societal: misinformation spread, erosion of trust in real expertise

Practice exercise

Ask AI: "What are the top 3 causes of death globally, with the most recent statistics?"

Then:
1. List each claim AI makes
2. Verify each against WHO data or a reputable source
3. Note any discrepancies
4. Did AI hallucinate? Did it use outdated numbers? Did it cite sources?

Recap

• AI hallucinates — confidently states false information
• Always verify: facts, citations, numbers, URLs, code, recent events
• Use the 3-source rule for important claims
• Use Perplexity for cited research, NotebookLM for your documents
• Watch for hallucination patterns: plausible but unverifiable, unsourced numbers, fake citations
• When in doubt, leave it out — or ask a human expert

Memory trick

V-FER — Verify Facts, Examples, References.

Or simpler: "If it matters, verify."

Last reviewed: 2026-08 · Skill 3 of 8

Identifying Hallucinations

Hallucinations are AI's biggest weakness. This skill teaches you to spot them before they cause harm.

What is a hallucination?

A hallucination is when AI generates information that is false, fabricated, or non-existent — but presents it with full confidence as if it were true.

Hallucinations are not bugs. They're a feature of how AI works: it predicts the next most likely token based on patterns. Sometimes the most likely-sounding token sequence is factually wrong.

Why hallucinations happen

AI doesn't "know" facts. It generates text that statistically resembles its training data. When you ask about:
- Something obscure, it generates plausible-sounding text
- Something recent, it extrapolates from old patterns
- Something specific (names, dates, citations), it sometimes invents

Types of hallucinations

1. Fabricated facts

AI states something that is simply not true.

Example: "The capital of Australia is Sydney" (it's Canberra)

2. Invented citations

AI generates realistic-looking academic citations that don't exist.

Example: "Smith, J. et al. (2023). Effects of caffeine on memory. Journal of Neuroscience, 45(3), 234-251." — paper doesn't exist

3. Non-existent URLs

AI generates URLs that look real but lead nowhere.

Example: "For more information, visit https://www.who.int/health-topics/caffeine-memory" — link doesn't exist

4. Misattributed quotes

AI attributes real quotes to the wrong person, or invents quotes for real people.

Example: "As Einstein said, 'Insanity is doing the same thing over and over and expecting different results.'" — Einstein didn't say this

5. Plausible but fake biographies

AI can generate convincing bios for non-existent people.

Example: Ask AI for a biography of "Dr. Maria Chen, inventor of the quantum stapler" — it will produce one

6. Hallucinated code

AI code can have subtle bugs, reference non-existent libraries, or use deprecated APIs.

7. False confidence on opinions

AI may state opinions as facts: "The best programming language is Python" without acknowledging subjectivity.

How to spot hallucinations

Red flags

• Specific numbers without sources — "73% of users prefer..."
• Exact quotes without citation — "As Steve Jobs said in 2007..."
• Confident claims about obscure topics
• Citations to specific papers, books, or court cases (always verify)
• Recent events (after training cutoff)
• URLs (always click before sharing)
• Code that uses libraries you don't recognize

The "too specific" test

If AI gives you extremely specific information (exact dates, exact numbers, exact quotes) on a topic that isn't widely documented, be suspicious. Specificity without sources is a hallucination red flag.

The "search test"

Copy the claim into Google. If no reputable source confirms it, it's likely hallucinated.

The "ask AI to verify" trick (use carefully)

You can ask AI: "Can you verify this claim with sources?"

But beware: AI may double down and generate fake verification. Always check the sources AI provides actually exist.

The hallucination spot-check

For any AI output, ask yourself:

1. Does this contain specific facts? If yes, verify them.
2. Does this contain citations? If yes, search for them.
3. Does this contain URLs? If yes, click them.
4. Does this contain numbers? If yes, find the original source.
5. Does this contain quotes? If yes, find the original source.
6. Does this sound too good to be true? If yes, be suspicious.
7. Is this a topic AI might not know much about? If yes, be extra suspicious.

Hallucination-prone topics

Be especially careful with:

• Medical information — AI can give dangerous advice
• Legal information — AI can invent laws or cases
• Financial advice — AI can give wrong investment guidance
• Historical events — AI can misstate dates, causes, or outcomes
• Recent events — anything after training cutoff
• Obscure topics — AI fills gaps with plausible fabrications
• Names and biographies — AI can invent people
• Code — AI can reference non-existent functions or libraries

How to reduce hallucinations

While you can't eliminate them, you can reduce them:

1. Use the right tool

• For current info: Perplexity (cites sources)
• For your documents: NotebookLM (only uses your docs)
• For math/code: Ask AI to use Python or show work

2. Add "I don't know" permission

Add to your prompt: "If you're not sure, say 'I don't know' rather than guessing."

This reduces hallucinations significantly.

3. Ask for sources upfront

Add: "For any factual claims, please cite your sources."

AI will either cite real sources or expose that it has none.

4. Use retrieval-augmented generation (RAG)

For important work, use tools that fetch real data before answering (Perplexity, NotebookLM, custom GPTs with web access).

5. Lower the temperature (for API users)

If you use AI via API, lower the temperature parameter (e.g., 0.3 instead of 0.7) for more deterministic, less creative responses.

What to do when you spot a hallucination

1. Don't trust the rest of the output — if AI hallucinated one thing, it may have hallucinated more
2. Re-ask the question with different phrasing
3. Use a different tool (Perplexity instead of ChatGPT)
4. Verify the corrected output the same way
5. Report it (some AI tools have feedback mechanisms)

Practice exercise

Test hallucination detection:

1. Ask ChatGPT: "Tell me about the 1987 Treaty of Lisbon" (there's no such treaty — Lisbon Treaty was 2007)
2. Note how confidently AI responds
3. Did it correct itself when challenged?
4. Try: "Tell me about Dr. James Hoffman, the quantum physicist who discovered the Hoffman effect" (no such person)
5. How convincing was the fake bio?

Recap

• Hallucinations are confident false information — AI's biggest weakness
• They happen because AI generates plausible text, not verified facts
• Watch for: specific numbers without sources, fake citations, non-existent URLs, misattributed quotes
• Spot-check: facts, citations, URLs, numbers, quotes
• Reduce hallucinations: use the right tool, add "I don't know" permission, ask for sources, use RAG
• For important claims: verify, verify, verify

Memory trick

SCUN — Sources, Citations, URLs, Numbers.

If AI gives you any of these, verify before trusting.

Last reviewed: 2026-08 · Skill 4 of 8

Protecting Your Privacy

What you share with AI tools may be stored, reviewed, or used to train future models. This skill teaches you what's safe to share and what isn't.

Why this skill matters

When you type something into ChatGPT, Claude, or Gemini, where does it go? Most people don't think about this. But the answer matters:

• Your prompts may be stored on company servers
• Your prompts may be reviewed by humans (for safety/quality)
• Your prompts may be used to train future AI models
• Your prompts may be leaked in a data breach
• Your prompts may be subpoenaed by law enforcement

This doesn't mean AI is evil. It means you should be intentional about what you share.

The golden rule

Never share anything with an AI tool that you wouldn't post on a public bulletin board.

If you wouldn't put it on a billboard with your name on it, don't put it in a chatbot.

What NEVER to share

Personal identifiers

• Social Security numbers / National ID numbers
• Passport numbers
• Driver's license numbers
• Bank account numbers
• Credit card numbers (full)
• Phone numbers (yours or others')
• Home addresses (yours or others')

Authentication data

• Passwords (yours or anyone's)
• PINs or security codes
• API keys or tokens
• Two-factor authentication codes
• Security question answers

Health information

• Medical records
• Mental health details
• Prescription information
• Genetic test results
• Anything that could be used to discriminate against you

Financial information

• Income details
• Tax returns
• Investment account balances
• Loan applications
• Credit reports

Confidential work information

• Proprietary company data
• Customer information
• Internal financials
• Unreleased product details
• Trade secrets
• Anything under NDA

Other people's information

• Other people's personal data (even their name + email)
• Other people's medical information
• Other people's financial information
• Photos of other people without consent

What's usually safe to share

• General questions about public topics
• Your own public information (LinkedIn, blog posts)
• Hypothetical scenarios (without real names/data)
• Public documents (news articles, published papers)
• General work questions (without proprietary details)
• Creative writing
• General advice requests

How to anonymize before sharing

If you need AI's help with something sensitive, anonymize first:

1. Replace names with placeholders

• "My coworker John" → "my coworker [A]"
• "Acme Corp" → "[Company]"
• "San Francisco" → "[City]"

2. Replace numbers with ranges

• "$87,432 salary" → "salary in the $80-100K range"
• "34 years old" → "in my 30s"

3. Remove identifying details

• Specific dates
• Specific locations
• Specific project names
• Specific client names

4. Use hypothetical framing

Instead of: "My boss Jane at Acme Corp said..."

Use: "If a manager at a mid-sized company said..."

Tool-specific privacy

ChatGPT (OpenAI)

• Free tier: Prompts may be used for training (you can opt out in settings)
• Plus/Pro: Same by default, opt out available
• Team/Enterprise: Not used for training by default
• Temporary Chat: New feature that doesn't save history

Claude (Anthropic)

• Free/Pro: Prompts may be used for training (opt out available)
• API: Not used for training
• Anthropic has a strong privacy stance, but always verify current policy

Gemini (Google)

• Free: Prompts may be reviewed by humans and used for training
• Google Workspace integration: Subject to your Workspace data policies
• API (Vertex AI): Not used for training

Perplexity

• Queries stored to improve service
• Cited sources are public (you're not revealing anything not already public)

NotebookLM

• Your uploaded documents are processed but not used to train models
• Relatively private — good for sensitive documents

How to opt out of training

ChatGPT

1. Settings → Data Controls
2. Toggle off "Improve the model for everyone"
3. Or use "Temporary Chat" for one-off conversations

Claude

1. Settings → Privacy
2. Opt out of training data usage

Gemini

1. Activity controls → Turn off Gemini Apps Activity
2. Note: this disables some features

Special concerns

For professionals

• Lawyers: Don't share client information; check your bar association's AI guidance
• Doctors: Don't share patient information (HIPAA)
• Therapists: Don't share client information (HIPAA)
• Accountants: Don't share client financial data
• HR: Don't share employee data

For parents

• Don't share your child's personal information
• Don't share photos of your child
• Don't share your child's school name, grade, or schedule

For job seekers

• Don't share your full SSN
• Don't share your full address
• Resume is generally OK (it's meant to be shared)

For business owners

• Don't share proprietary data
• Don't share customer information
• Don't share financial details
• Don't share employee information

The privacy checklist

Before hitting "send" on any AI prompt, ask:

• [ ] Did I include any personal identifiers? (SSN, ID, account numbers)
• [ ] Did I include any authentication data? (passwords, keys, codes)
• [ ] Did I include any health information?
• [ ] Did I include any financial information?
• [ ] Did I include any confidential work information?
• [ ] Did I include other people's information without consent?
• [ ] Would I be comfortable if this prompt appeared in a data breach?

If you checked any box, anonymize or remove that information before sending.

What to do if you accidentally shared sensitive info

1. Delete the conversation (most tools allow this)
2. Change compromised passwords immediately
3. Monitor for misuse (credit reports, account activity)
4. Contact the AI provider to request deletion
5. For serious breaches (SSN, financial), contact relevant authorities

Recap

• Never share anything with AI you wouldn't post publicly
• Never share: personal identifiers, auth data, health info, financial info, confidential work data, others' info
• Anonymize before sharing sensitive scenarios
• Check each tool's privacy policy and opt out of training
• Use the privacy checklist before every prompt
• If you accidentally share sensitive info, act fast to mitigate

Memory trick

PHAFECO — Personal, Health, Auth, Financial, Employment, Confidential, Others'.

If it fits any of these categories, don't share it.

Last reviewed: 2026-08 · Skill 5 of 8

Using AI Responsibly

Ethics isn't abstract — it's the daily choices you make about how to use AI. This skill teaches you the principles.

Why this skill matters

Every time you use AI, you're making ethical choices:

• Do I disclose that I used AI?
• Do I verify before sharing?
• Do I use AI to deceive or to assist?
• Do I respect copyright and creators' rights?
• Do I avoid harm to others?

These choices compound. Millions of people making responsible choices → AI benefits society. Millions making irresponsible choices → AI harms society.

The 6 principles of responsible AI use

1. Verify before trusting

We covered this in Skill 3: Fact-Checking. Verify facts, citations, URLs, and numbers before acting on them or sharing them.

2. Disclose AI use

When AI assistance is non-trivial, disclose it:

• Academic work: Cite AI tools used, following your institution's policy
• Professional work: Tell your team and clients when AI helped substantially
• Creative work: Be transparent with your audience
• Decisions affecting others: Always disclose (hiring, lending, medical)

Rule of thumb: If removing AI would change the work significantly, disclose.

3. Don't deceive

AI makes deception easy. Don't:

• Generate fake reviews (positive or negative)
• Create deepfakes of real people without consent
• Pass AI-generated content as human-created when context expects human creation
• Impersonate others using AI voice/video cloning
• Generate fake testimonials or social proof
• Use AI to manipulate or scam

4. Respect copyright

• Don't ask AI to "write in the style of [living author]" and publish it
• Don't use AI to reproduce copyrighted works
• Don't generate content that infringes trademarks
• When in doubt, create original work or get permission

5. Avoid harm

• Don't use AI to harass, bully, or threaten
• Don't generate content that promotes self-harm, violence, or illegal acts
• Don't use AI to discriminate (hiring, lending, housing)
• Don't generate content that sexualizes minors
• Don't use AI to spread misinformation

6. Don't delegate judgment

AI can help you think, but it can't make decisions for you. Especially for:

• Hiring/firing decisions: AI can help screen, but humans decide
• Medical decisions: AI can inform, but doctors decide
• Legal decisions: AI can research, but lawyers decide
• Financial decisions: AI can analyze, but you decide
• Parenting decisions: AI can suggest, but you decide

Specific ethical scenarios

Scenario 1: Student using AI for homework

Responsible: Use AI to brainstorm, explain concepts, quiz yourself, get feedback on drafts. Cite AI use per your school's policy.

Irresponsible: Submit AI-generated text as your own. Use AI to write essays you didn't write. Use AI to solve problems you should learn to solve yourself.

Scenario 2: Professional writing emails

Responsible: Use AI to draft, edit, and improve your emails. The content is yours; AI is a tool.

Irresponsible: Let AI write all your emails without reading them. Send AI-generated content that's wrong because you didn't verify.

Scenario 3: Content creator using AI

Responsible: Use AI to brainstorm, outline, draft, edit. Disclose AI assistance to your audience. Create original work.

Irresponsible: Pass AI-generated content as fully human. Generate content in the style of specific creators without credit. Use AI to mass-produce low-quality content.

Scenario 4: Hiring manager using AI

Responsible: Use AI to write job descriptions, summarize resumes, prepare interview questions. Human makes all hiring decisions.

Irresponsible: Use AI to filter candidates by demographic indicators. Use AI to make hiring decisions. Use AI to analyze video interviews for "personality" without consent.

Scenario 5: Using AI image generators

Responsible: Use for original creative work. Use commercially-safe tools (Adobe Firefly) for commercial projects. Respect content filters.

Irresponsible: Generate deepfakes of real people. Generate content that infringes copyright. Use AI images without disclosure in journalism.

The responsibility checklist

Before using AI output, ask:

• [ ] Did I verify the facts?
• [ ] Did I disclose AI use where expected?
• [ ] Am I being honest about what AI did vs. what I did?
• [ ] Did I respect copyright and others' rights?
• [ ] Am I avoiding harm to others?
• [ ] Did I make the final judgment, not AI?
• [ ] Would I be comfortable if everyone knew how I used AI?

When in doubt

If you're unsure whether a use is ethical:

1. Don't do it until you've thought it through
2. Ask: "Would I be comfortable if this were on the front page of a newspaper?"
3. Consult: For work, ask your manager or ethics team. For school, ask your teacher.
4. Err on the side of disclosure: When in doubt, disclose.

The long view

Every individual ethical choice seems small. But collectively, they determine whether AI becomes a force for good or harm. Your choices matter.

Recap

• 6 principles: verify, disclose, don't deceive, respect copyright, avoid harm, don't delegate judgment
• Use the responsibility checklist before sharing/acting on AI output
• When in doubt, disclose
• Your choices compound — responsible AI use is a daily practice

Memory trick

VD-DR-AH-DJ — Verify, Disclose, Don't deceive, Respect copyright, Avoid harm, Don't delegate Judgment.

Or simpler: "Would I be comfortable if everyone knew?"

Last reviewed: 2026-08 · Skill 6 of 8

Writing Effective Prompts

Building on "Asking Better Questions" — the specific craft of writing prompts that consistently get great AI outputs.

Why this skill matters

A prompt is just a question with structure. Once you know the structure, you can apply it to any task — writing, coding, analysis, brainstorming, anything.

The prompt formula

Every great prompt has 5 parts. Memorize this:

[ROLE] + [CONTEXT] + [TASK] + [FORMAT] + [CONSTRAINTS]

1. ROLE — who AI should be

Tell AI what perspective to take. This dramatically affects output.

Examples:
- "You are an experienced marketing copywriter"
- "You are a high school teacher explaining to a struggling student"
- "You are a senior software engineer doing code review"
- "You are a friend who's honest but kind"

2. CONTEXT — the situation

What does AI need to know to give you a useful answer?

Examples:
- "I'm a freelance designer with 3 years of experience, pitching to a small tech startup"
- "The audience is 50 middle managers at a manufacturing company, most of whom have never used AI"
- "This is for a 5-minute lightning talk, not a 60-minute workshop"

3. TASK — what specifically to do

One clear task per prompt. If you have 3 tasks, send 3 prompts.

Examples:
- "Write a 200-word cold outreach email"
- "Summarize this article in 5 bullet points"
- "Explain quantum computing at 3 levels: 10-year-old, 15-year-old, college student"

4. FORMAT — how to structure the output

Examples:
- "Format as a numbered list"
- "Use a table with columns: Tool, Price, Best for"
- "Write in markdown with H2 headers for each section"
- "Output as JSON with these keys: title, summary, tags"

5. CONSTRAINTS — what to avoid or limit

Examples:
- "Under 300 words"
- "Avoid jargon"
- "Don't use the words 'revolutionary' or 'game-changing'"
- "Tone: warm but professional, not corporate"

A complete example

Bad prompt:

"Write a blog post about AI for students"

Good prompt (using the formula):

"You are an experienced education writer who's written for The Atlantic and Edutopia. I'm a college student starting a blog about AI for students. Write a 600-word blog post titled '5 Ways AI Can Help You Study (Without Doing the Work for You)'. Target audience: undergrads who are curious about AI but worried about academic integrity. Format: intro paragraph, 5 numbered sections with H2 headers, conclusion. Tone: conversational, specific, no hype. Avoid: bullet-point-only sections (each section needs prose), clichés like 'in today's fast-paced world'."

Advanced techniques

Chain of thought

For complex reasoning, ask AI to think step by step:

"Before answering, walk through your reasoning step by step. Then give your final answer."

This significantly improves accuracy on math, logic, and analysis tasks.

Few-shot examples

Show AI 2-3 examples of what you want:

"Here are 3 examples of tweets I like: [paste 3]. Now write 5 more in the same style about [topic]."

Iterative refinement

Don't expect perfection on the first try. Use a conversation:

1. First prompt: get a draft
2. Second prompt: "Make it shorter and more punchy"
3. Third prompt: "The second paragraph is weak — rewrite it with a specific example"
4. Fourth prompt: "Now give me 2 alternative openings"

Ask for alternatives

"Give me 3 different versions: one formal, one casual, one witty. I'll pick the best."

This is often faster than iterating on one version.

Use AI to critique itself

"Here's what you wrote: [paste]. Now critique it as a harsh editor would. What's weak? What should be cut?"

Common prompt mistakes

• Too many tasks in one prompt — split them up
• No format specified — you'll get walls of text
• No length constraint — you'll get either too short or too long
• Vague role — "be helpful" is not a role
• No examples — AI guesses what you want, often wrong

The prompt library

You don't need to write every prompt from scratch. See our Prompt Library for 200+ copy-paste prompts across 11 audiences.

Practice exercise

Rewrite this weak prompt using the 5-part formula:

"Help me write a cover letter"

Your improved version should include role, context, task, format, and constraints. Compare to the example in src/prompts/for-job-seekers.md.

Recap

• Use the formula: [ROLE] + [CONTEXT] + [TASK] + [FORMAT] + [CONSTRAINTS]
• One task per prompt
• Always specify format and length
• Use chain-of-thought for complex reasoning
• Use few-shot examples to show what you want
• Iterate — don't expect perfection on try 1
• Ask for alternatives, not just one answer
• Use AI to critique itself

Memory trick

RCFTC — Role, Context, Format, Task, Constraints.

Or: "Who am I talking to, what's the situation, what do I want, how should it look, what should I avoid?"

Last reviewed: 2026-08 · Skill 2 of 8