This guide serves two purposes. As a learning tool, read Parts I and II in order — each chapter builds on the last. As a reference document, jump straight to the Prompt Pattern Library (Chapter 6), the Accuracy Checklist (Chapter 12), or the Glossary (Chapter 13) whenever you need them.

Every prompt template in this guide is ready to copy and use. Words in [BRACKETS] are placeholders — replace them with your own topic. Example: "Explain [TOPIC] in plain English" becomes "Explain Medicare Part D in plain English."

---

The one-sentence version

Tools like Claude and ChatGPT are prediction machines: they have read enormous amounts of text, and they answer you by predicting, one word at a time, what a knowledgeable response would look like.

The library analogy

Imagine someone who spent years reading nearly everything — books, articles, websites, manuals — but was not allowed to keep any of it. They can't quote page 200 of a specific book, yet they absorbed the patterns in everything they read: how doctors explain diagnoses, how lawyers structure arguments, how recipes are written.

That's an AI language model. It doesn't look things up in a database when you ask a question. It reconstructs an answer from patterns it learned. This explains both its strengths and weaknesses:

• Strength: It can explain, summarize, compare, translate, brainstorm, and draft about almost any subject, instantly.
• Weakness: Because it reconstructs rather than retrieves, it can produce something that sounds right but isn't — like a person misremembering a fact with total confidence. This is called a hallucination, and Chapter 3 covers it in depth.

Key terms you'll hear constantly

AI (Artificial Intelligence) — the broad field of making computers do things that normally require human intelligence.

LLM (Large Language Model) — the specific kind of AI behind Claude and ChatGPT. "Large" refers to being trained on vast amounts of text.

Chatbot / Assistant — the product you talk to. Claude and ChatGPT are chat interfaces built on top of LLMs.

Prompt — whatever you type to the AI. Improving your prompts is the single highest-value skill you can learn, and it's the core of this guide.

Training data — the text the model learned from. It has a cutoff date, which is why models may not know recent events unless they search the web.

Token — the chunks of text (roughly word pieces) the AI reads and writes. You mostly don't need to think about tokens, but long conversations can exceed the model's memory (its "context window"), which is why an AI can lose track of things said much earlier.

Context window — the AI's short-term memory: everything in the current conversation it can "see." Start a new chat and that memory is wiped clean.

What AI is genuinely good at

1. Explaining — any concept, at any level of simplicity, with analogies.
2. Summarizing — condensing long documents into key points.
3. Drafting — letters, emails, plans, outlines you then refine.
4. Comparing — laying out options side by side.
5. Brainstorming — generating possibilities you hadn't considered.
6. Transforming — turning notes into prose, prose into bullet points, formal into casual.
7. Teaching — acting as a patient tutor that never tires of questions.

What AI is unreliable at (without help)

1. Precise facts — dates, statistics, quotations, citations, prices. Always verify.
2. Recent events — anything after its training cutoff, unless it searches the web.
3. Math on large numbers — it predicts text; it doesn't naturally calculate (though modern tools can run code to compute accurately when asked).
4. Knowing what it doesn't know — it rarely says "I'm not sure" unless you invite it to.
5. Consistency — ask the same question twice and you may get different answers.

The rest of this guide teaches you to get the good and defend against the bad.

---

You use both Claude (made by Anthropic) and ChatGPT (made by OpenAI). They are more alike than different: both are LLM chat assistants, both work on iPad, iPhone, and PC, and every prompt technique in this guide works on both.

What they have in common

• Free tiers plus paid subscriptions that unlock stronger models and higher usage limits.
• Apps for iPhone/iPad plus browser access on your HP computer (claude.ai and chatgpt.com).
• The ability to upload documents and images and ask questions about them.
• Web search for current information (availability varies by plan and settings).
• Conversation history you can return to.

Practical differences (as a user, not a technician)

• Claude is often praised for careful, nuanced writing, long-document analysis, and honest handling of uncertainty. Claude "Projects" let you store standing instructions and reference files.
• ChatGPT has a broad feature set — image generation, voice conversation modes, and "custom GPTs." Its "memory" feature can carry facts about you between chats (you can turn this off).
• Both companies update features frequently, so treat any specific feature comparison as a snapshot. The skills in this guide are durable; the feature lists are not.

A smart two-tool habit

Because both tools can hallucinate, owning two is an advantage: ask the same important question in both, and compare. If Claude and ChatGPT independently agree, confidence rises. If they disagree, you've found exactly the point that needs real verification. Chapter 8 turns this into a formal technique.

Hygiene habits worth building from day one

1. One topic per conversation. Starting a fresh chat for a new subject keeps the AI focused and avoids confusion from leftover context.
2. Name your chats (or let the app name them) so you can find them later.
3. Don't share sensitive personal data — Social Security numbers, passwords, full account numbers — in any chat.
4. Treat answers as drafts, not verdicts, for anything involving money, health, or law.

---

What a hallucination is

A hallucination is when an AI states something false as if it were true — an invented statistic, a citation to a study that doesn't exist, a confident wrong date, a fabricated quote. The AI isn't lying; lying requires knowing the truth. It's doing exactly what it was built to do — produce plausible text — in a situation where plausible and true have come apart.

Why it happens

1. Prediction, not lookup. The model generates what an answer typically looks like. If real knowledge is thin, the pattern gets filled in with plausible invention.
2. Trained to be helpful. Models learn that complete, confident answers are rated well. "I don't know" was historically underrewarded — so you must explicitly make uncertainty welcome.
3. No built-in fact-checker. Nothing inside the model automatically compares its output to reality.
4. Your question can plant errors. Ask "Why is X true?" about something false, and the AI often plays along. This is called a leading question problem.

Where hallucinations cluster — the risk map

High risk: specific citations and references; exact numbers and statistics; quotes attributed to real people; details about obscure people, places, or products; recent events; legal and medical specifics; anything you asked in a leading way.

Lower risk: broad concepts, well-documented general knowledge, explanations of common processes, summarizing a document you provided (it can see the actual text).

The five warning signs of a weak answer

1. Suspicious precision — an exact figure ("73.4% of retirees…") with no source.
2. Uniform confidence — every claim delivered with equal certainty; real experts vary their confidence.
3. Too-perfect examples — a study, case, or quote that fits your question exactly.
4. Vague sourcing — "studies show," "experts agree," "it is widely known."
5. Agreement with your slant — the answer mirrors the assumption in your question instead of examining it.

Your first three defenses (the full system is Chapter 7)

1. Add to important prompts: "If you are not certain, say so. Do not guess."
2. Ask afterward: "State your confidence in each claim above: high, medium, or low."
3. For anything that matters: verify independently — a second AI, a web search, or a primary source.

A calibration exercise

Try this once in each tool: ask about something you know deeply — your profession, hometown, a hobby. Notice what it gets right, what it flubs, and how confident it sounds while flubbing. That feeling — "it sounds authoritative but I can see the errors" — is the healthy skepticism to carry into topics you don't know.

---# PART II: PROMPTING SKILLS

The core insight

The quality of the answer is set by the quality of the question. A vague prompt gets a generic answer; a specific prompt gets a specific one. You are not "searching" — you are briefing an assistant.

The five building blocks: R-T-C-F-C

Strong prompts tend to contain five elements. You won't need all five every time, but knowing them lets you diagnose why a prompt underperformed.

1. Role — who the AI should be.

"Act as a patient financial educator."

2. Task — the verb. What exactly to do: explain, compare, summarize, list, draft, critique, analyze.

"Compare the two Medicare supplement plans below."

3. Context — what it needs to know about your situation.

"I am 65, retired, in Ohio, and my priority is low out-of-pocket risk."

4. Format — what the answer should look like.

"Give me a table with a plain-English row for each difference, then a 3-sentence recommendation."

5. Constraints — rules and guardrails.

"Use no jargon. If information may be outdated, say so. Do not guess at prices."

Weak vs. strong — one example

Weak: "Tell me about annuities."

Strong: "Act as a retirement-income educator. Explain fixed annuities to a 65-year-old novice: what they are, the top 3 benefits, top 3 risks, and the questions I should ask a salesperson before buying. Use plain English and everyday analogies. If any point depends on current rates or state rules, flag it as 'verify locally' rather than guessing."

Same tool, same effort to type — dramatically different value.

Seven habits of effective prompters

1. Be specific about your situation. Age, location, goal, and constraints change the right answer.
2. Say who the answer is for. "Explain to a complete beginner" vs. "assume I know the basics."
3. Ask for structure. Tables, numbered steps, and pros/cons lists are easier to use than essays.
4. Set the length. "In 200 words" or "in as much detail as needed."
5. Iterate. The first answer is a starting point. Reply with "simpler," "shorter," "more detail on point 2," "give me an example."
6. Ask for the question behind your question. "What should I have asked that I didn't?" is consistently one of the highest-value follow-ups.
7. Invite pushback. "Challenge my assumptions" and "what am I missing?" counteract the AI's tendency to agree with you.

The follow-up toolkit

After any answer, these six replies do most of the work:

• "Explain that more simply."
• "Give me a concrete example."
• "What are the counterarguments?"
• "What's the weakest part of that answer?"
• "What would an expert add?"
• "Summarize this whole conversation in 10 bullet points." (great before ending a long chat)

---

These are short additions that measurably change answer quality. They work because each one counteracts a specific AI weakness. Grouped by what they fix.

Group A — Force honesty about uncertainty (fixes overconfidence)

1. "If you are not certain, say so — do not guess." The single most valuable sentence in this guide.
2. "State your confidence level (high/medium/low) for each claim." Turns uniform confidence into a map of what to verify.
3. "Distinguish between what is well-established and what is debated." Separates textbook fact from contested territory.
4. "What are the limits of your knowledge on this?" Surfaces cutoff dates and thin areas.
5. "Only include information you can support; omit anything speculative."
6. "Flag anything that may have changed recently." Catches training-cutoff problems.

Group B — Force evidence and sourcing (fixes vague authority)

1. "Cite the type of source for each claim (study, law, official guidance, common knowledge)."
2. "What evidence supports this? What evidence contradicts it?"
3. "Separate facts from assumptions." A workhorse for any analysis.
4. "Identify every unsupported claim in your answer." Ask this after an answer — the AI audits itself surprisingly well.
5. "What would change your conclusion?" Reveals how solid the reasoning is.
6. "Base your answer only on the document I provided." Essential when analyzing your own files — stops outside "knowledge" from leaking in.

Group C — Improve reasoning depth (fixes shallow answers)

1. "Think step by step." Classic; produces visible reasoning you can check.
2. "Consider multiple perspectives before concluding."
3. "Play devil's advocate against your own answer."
4. "What is the strongest case for the opposite view?"
5. "First principles: explain this from the ground up without relying on convention."
6. "Before answering, list what information you'd ideally need — then note which of it you actually have."

Group D — Control scope and format (fixes rambling)

1. "In plain English, no jargon."
2. "Be concise — no filler, no preamble."
3. "Answer in a table." / "Answer as numbered steps."
4. "Give me the 80/20: the 20% of this topic that delivers 80% of the value."
5. "Limit your answer to what a beginner needs to know first."
6. "Bottom line first, then supporting detail."

Group E — Expose blind spots (fixes agreement bias)

1. "What am I not asking that I should be?"
2. "Challenge the assumptions in my question."
3. "What do people commonly get wrong about this?"
4. "What are the risks or downsides you haven't mentioned?"
5. "If this answer were wrong, what would be the most likely reason?"
6. "Steelman both sides before giving your view." ("Steelman" = present each side at its strongest.)

How to combine them

Don't stack all 30. Pick one from Group A (honesty) plus whichever group matches your task. A practical everyday combo to memorize:

"[Your question]. Be concise and use plain English. Separate facts from assumptions, state your confidence, and if you're not certain about something, say so rather than guessing."

---

This is the heart of the reference guide: 100 field-tested prompt patterns in 10 categories. Copy the template, replace the [BRACKETS], and go. Each works in both Claude and ChatGPT.

Category 1: Fact-Finding (Patterns 1–10)

1. Direct fact with honesty guard

What is [FACT YOU NEED]? If you are not certain or this may have changed since your training data, say so explicitly rather than guessing.

2. The known/unknown split

Regarding [TOPIC]: list what is definitively known, what is believed but uncertain, and what is unknown or disputed.

3. Date-sensitive fact check

As of your most recent knowledge: [QUESTION]. State the date your knowledge likely ends for this topic and what kind of changes may have occurred since.

4. Number with provenance

What is [STATISTIC]? Tell me the type of source it would come from, the year of the estimate, and how much such figures typically vary between sources.

5. Definition triangulation

Define [TERM] three ways: (1) plain English for a beginner, (2) the formal/technical definition, (3) how it's used loosely in everyday conversation. Note where the loose usage misleads.

6. Who-what-when-where scaffold

Give me the essential facts about [EVENT/THING]: who, what, when, where, why it matters. One line each. Flag any element you are unsure about.

7. The common misconception filter

What do most people get wrong about [TOPIC]? For each misconception, give the correction and why the myth persists.

8. Boundaries of a rule

Explain [RULE/LAW/GUIDELINE] and — more importantly — its exceptions, limits, and the situations where people wrongly assume it applies.

9. Both-tools cross-check (use in each AI)

[QUESTION]. Answer carefully. I will be comparing your answer against another source, so flag any point you are less than fully confident about.

10. The "verify locally" separator

[QUESTION ABOUT RULES, PRICES, OR SERVICES]. Separate your answer into (a) general principles that are stable and (b) specifics that vary by location/provider/time, which I should verify locally. Do not guess at the specifics.

Category 2: Research (Patterns 11–20)

11. Research briefing

Act as a research analyst. Give me a briefing on [TOPIC]: background, current state, key players or factors, main open questions, and what a careful person should read or check next. Separate established facts from interpretation.

12. The landscape map

Map the landscape of [TOPIC]: the main schools of thought or approaches, who holds each view, and the strongest evidence for each. Do not pick a winner yet.

13. Structured deep dive

I want to research [TOPIC] thoroughly. First, give me an outline of the 5–8 sub-questions I need to answer to truly understand it. Then we'll take them one at a time.

14. Source-type shopping list

For researching [TOPIC], what are the best types of sources (government data, peer-reviewed studies, industry reports, primary documents), what is each good for, and what are the specific top sources of each type I should look for?

15. The 80/20 overview

Give me the 80/20 on [TOPIC]: the 20% of the information that provides 80% of the understanding. Plain English, no filler.

16. History-to-present arc

Trace [TOPIC] from its origins to today in 5–8 key turning points. For each: what changed, why it mattered, and what misconception it left behind.

17. Data interpretation guard

Here is a statistic I found: [STATISTIC + SOURCE]. Help me interpret it carefully: what it does and does not show, common misreadings, what context is missing, and what I'd need to know before relying on it.

18. The opposing-expert simulation

Simulate a respectful debate between two well-informed experts who disagree about [TOPIC]. Have each make their 3 strongest points and respond to the other. Then summarize where the genuine uncertainty lies.

19. Literature-style review

Summarize the state of evidence on [QUESTION] the way a review article would: what most studies find, what dissenting studies find, methodological weaknesses on each side, and the honest bottom line including uncertainty.

20. Research session closer

Summarize everything we've established in this conversation as a research memo: findings, confidence level for each, open questions, and recommended next steps. Format for saving to my files.

Category 3: Learning & Teaching (Patterns 21–30)

21. The patient tutor

Act as a patient tutor teaching me [TOPIC] from zero. Start with the big picture in plain English, then check my understanding with one question before going deeper. Never move on until I confirm I've got it.

22. Explain like I'm smart but new

Explain [TOPIC] to someone intelligent but completely new to the field. Use analogies to everyday life. Define every technical term the first time you use it.

23. The analogy generator

Explain [CONCEPT] using three different analogies: one from cooking, one from home/household life, one from driving or travel. Then note where each analogy breaks down.

24. Progressive depth ladder

Explain [TOPIC] at three levels: (1) one sentence a child could understand, (2) one paragraph for a curious adult, (3) one page for someone who wants real depth. Label each level.

25. The learning plan

Design a self-study plan for me to learn [TOPIC] in [TIMEFRAME], spending about [TIME] per week. Break it into weekly milestones with what to learn, how to practice it, and how I'll know I've learned it.

26. Quiz me

Quiz me on [TOPIC] with 10 questions, one at a time, starting easy and getting harder. After each answer I give, tell me if I'm right, explain why, and adjust difficulty. Keep score.

27. The Feynman test

I'm going to explain [TOPIC] back to you in my own words. Point out every place where my understanding is wrong, fuzzy, or incomplete — be direct, not polite. Here's my explanation: [YOUR EXPLANATION]

28. Mistake-driven learning

What are the 5 most common mistakes beginners make when learning [TOPIC or SKILL], why does each happen, and what habit prevents it?

29. Connect to what I know

I already understand [FAMILIAR TOPIC] well. Teach me [NEW TOPIC] by building on that knowledge — map the concepts I know onto the new ones, and flag where the mapping misleads.

30. The spaced-review generator

Based on what you've taught me about [TOPIC] in this conversation, create a one-page review sheet: key concepts, memory hooks, and 5 self-test questions with answers at the bottom.

Category 4: Summarization (Patterns 31–40)

31. Layered summary

Summarize the following in three layers: a one-sentence essence, a 5-bullet key-points summary, and a one-page detailed summary. Base it ONLY on the text I provide, not outside knowledge. [PASTE TEXT]

32. The action extract

From the document below, extract only: (1) decisions made, (2) actions required and by whom, (3) deadlines, (4) open questions. Ignore everything else. [PASTE TEXT]

33. Summarize for a purpose

Summarize this document specifically to help me [YOUR PURPOSE — e.g., decide whether to sign it]. Emphasize what matters for that purpose and flag anything concerning. [PASTE TEXT]

34. The skeptical summary

Summarize the argument in this text, then critique it: what claims lack support, what's exaggerated, what's omitted, and what the other side would say. [PASTE TEXT]

35. Plain-English translation

Rewrite the following in plain English at a general-audience reading level, keeping all substantive meaning. Flag any passage whose meaning is ambiguous in the original. [PASTE TEXT]

36. Compare-the-coverage

Here are two accounts of the same matter. Summarize where they agree, where they disagree, and what each includes that the other omits. [PASTE BOTH]

37. The meeting/conversation digest

Turn these rough notes into a clean summary: main points discussed, decisions, disagreements, and follow-ups. Do not invent details that aren't in the notes. [PASTE NOTES]

38. Long-document interrogation

I've uploaded a long document. Before summarizing, list the 8–10 questions this document answers. I'll pick which ones I care about, and you'll answer only those, citing the section each answer comes from.

39. The "what's buried" scan

Review this document for anything a careless reader would miss but shouldn't: fine print, exceptions, fees, obligations, deadlines, or unusual terms. Quote each one exactly. [PASTE TEXT]

40. Summary integrity check

Here is a summary and the original. Check the summary against the original: list anything the summary gets wrong, overstates, or omits that matters. [PASTE BOTH]

Category 5: Comparison (Patterns 41–50)

41. The decision table

Compare [OPTION A] vs [OPTION B] for someone in my situation: [YOUR SITUATION]. Build a table of the factors that matter most to me, then give a bottom line with your confidence level.

42. Criteria-first comparison

Before comparing [OPTIONS], first tell me: what criteria SHOULD someone in my situation use to judge these? Let me confirm the criteria, then compare against them.

43. The tie-breaker

I'm torn between [A] and [B]. Argue the full case for A, then the full case for B, then tell me what single piece of information would break the tie.

44. Apples-to-apples enforcement

Compare [A] and [B], but first normalize them: state each one's assumptions, units, and included/excluded costs so the comparison is truly apples-to-apples. Flag where sellers make them hard to compare on purpose.

45. The switching-costs lens

Beyond comparing [A] vs [B] directly, compare the cost of being WRONG with each: what happens if I choose A and regret it vs choose B and regret it? Which mistake is easier to undo?

46. Three-option escape

I'm comparing [A] and [B]. What option C (or a hybrid, or "neither/wait") am I not considering? Evaluate it against the same criteria.

47. Feature-vs-need mapping

Here are the features of [PRODUCT/SERVICE OPTIONS]. Here is my actual situation: [SITUATION]. Map features to my real needs, and mark which advertised features I would genuinely never use.

48. The expert-buyer simulation

If a seasoned expert in [DOMAIN] were choosing between [A] and [B] for a family member, what would they check that a novice wouldn't? What questions would they ask the seller?

49. Total-cost comparison

Compare the TRUE total cost of [A] vs [B] over [TIME PERIOD]: upfront, recurring, hidden fees, likely maintenance, and exit costs. State which figures are estimates and how uncertain they are.

50. Side-by-side claims audit

Both of these sources make competing claims: [PASTE/DESCRIBE]. Build a claim-by-claim table: what each asserts, what evidence each offers, and which claims are actually in conflict vs just emphasized differently.

Category 6: Decision-Making (Patterns 51–60)

51. The decision brief

Help me decide: [DECISION]. My situation: [CONTEXT]. My priorities, in order: [PRIORITIES]. Give me a decision brief: options, how each scores against my priorities, key risks, and a recommendation with confidence level. Challenge my priorities if they seem off.

52. Pre-mortem

Assume I made [DECISION] and one year later it turned out badly. Write the story of what went wrong — the 3 most plausible failure paths. Then tell me what I could check now to guard against each.

53. The reversibility sort

For decision [DECISION], sort the moving parts into reversible vs irreversible. Where can I keep options open, delay commitment, or run a small test before deciding fully?

54. Decision matrix builder

Build a weighted decision matrix for [DECISION]. Propose criteria and weights based on my situation: [SITUATION]. Show the scoring for each option, then tell me how sensitive the result is — would a small change in weights flip the answer?

55. The 10/10/10 test

For [DECISION], how will I likely feel about each option in 10 days, 10 months, and 10 years? What does that lens reveal that the pros/cons list hides?

56. Assumption smoke-out

Here's my current plan: [PLAN]. List every assumption this plan silently relies on, rank them by how damaging it would be if wrong, and suggest how to test the top three cheaply.

57. The trusted-advisor panel

Convene a panel on my decision [DECISION]: a cautious financial advisor, an experienced person who has done this before, and a skeptical consumer advocate. Each gives their honest take, including where they disagree with each other.

58. Minimum information to decide

For [DECISION], what is the smallest set of facts I need before deciding is responsible? Which of those do I already have ([LIST WHAT YOU KNOW]) and how do I get the rest?

59. The default-option challenge

My default plan is to [DEFAULT]. Argue against the default: who benefits from me choosing it, what does inertia cost me, and what would have to be true for the default to genuinely be best?

60. Decision journal entry

Turn our discussion of [DECISION] into a decision-journal entry: the decision, date, my reasoning, key assumptions, what I predict will happen, and what would tell me I was wrong. (I'll save this to review later.)

Category 7: Root-Cause Analysis (Patterns 61–70)

61. Five Whys

Use the Five Whys technique on this problem: [PROBLEM]. Ask "why" iteratively to trace surface symptoms to root causes. Show each step, and note where the chain could branch in more than one direction.

62. Symptom vs cause sorter

Here's everything I'm observing: [LIST]. Sort these into symptoms, contributing factors, and candidate root causes. What test or question would distinguish between the candidate causes?

63. The differential diagnosis

Like a diagnostician: given [PROBLEM/SYMPTOMS], list the possible causes from most to least likely, what evidence supports each, and what I could check to rule each one in or out.

64. Timeline-to-cause

Here's the sequence of events: [TIMELINE]. What changed right before the problem appeared? Identify candidate triggers and distinguish correlation from causation for each.

65. The fishbone sweep

Analyze [PROBLEM] across all major cause categories (people, process, equipment/tools, environment, materials, methods). Ask me questions where you need more information rather than assuming.

66. Recurrence detector

This problem keeps coming back: [PROBLEM + HISTORY]. Why do the fixes keep failing? Distinguish between the root cause and the reason the fixes don't stick — they're often different.

67. The "what would have prevented this" lens

Given what happened ([EVENT]), work backward: at what points could this have been prevented, what would prevention have required at each point, and which prevention is most practical going forward?

68. Human-error deep look

The easy answer is "someone made a mistake": [SITUATION]. Go deeper — what about the situation, incentives, information, or design made that mistake easy to make? Assume the person acted reasonably given what they knew.

69. Contradiction hunter

Something doesn't add up: [DESCRIBE THE INCONSISTENCY]. List the possible explanations for the contradiction, including mundane ones (typos, timing, different definitions) before dramatic ones. How would I test each?

70. Root-cause report

Summarize our root-cause analysis of [PROBLEM] as a one-page report: symptom, investigation steps, root cause(s) with confidence level, contributing factors, and recommended corrective actions ranked by impact and ease.

Category 8: Risk Assessment (Patterns 71–80)

71. Risk register

Build a risk register for [PLAN/SITUATION]: each significant risk, its likelihood (low/med/high), impact if it happens, early warning signs, and what I can do to reduce or prepare. Sort by likelihood × impact.

72. The worst-case ladder

For [SITUATION/DECISION]: what's the realistic worst case, the plausible bad case, the expected case, and the good case? What drives the difference between them?

73. Scam and manipulation screen

Here's an offer/message I received: [PASTE OR DESCRIBE]. Analyze it for signs of scams or manipulation: pressure tactics, too-good-to-be-true elements, requests for unusual payment or information, impersonation signs. What would a fraud investigator check?

74. The fine-print risk scan

Review this agreement/policy for the risks it shifts onto me: [PASTE TEXT]. What can the other party change unilaterally? What am I agreeing to that isn't obvious? What's missing that should be there?

75. Single point of failure finder

In my plan/setup ([DESCRIBE]), find the single points of failure — the things that, if they break, break everything. How exposed am I, and what's the cheapest redundancy for each?

76. The dependency check

For [PLAN] to work, what has to go right that is outside my control? Rank those external dependencies by fragility and tell me which ones I should have a fallback for.

77. Blind-spot sweep

Given my description of [SITUATION], what risks am I probably not seeing because of how I've framed it? What would someone with the opposite worry notice first?

78. Safety margin calculator

My plan assumes [KEY ASSUMPTIONS/NUMBERS]. Stress-test it: what happens if costs run 25% higher, timelines 50% longer, or returns 30% lower? Where does the plan break first?

79. The irreversibility flag

In [PLAN/DECISION], flag every step that is expensive or impossible to undo. For each: what should I be certain of BEFORE that step, and can the step be delayed, tested, or made reversible?

80. Risk communication check

I read that [RISK CLAIM, e.g., "X doubles your risk of Y"]. Put this in perspective: absolute vs relative risk, the baseline rate, who the finding applies to, and whether the size of the risk justifies the alarm of the headline.

Category 9: Timeline & Chronology (Patterns 81–90)

81. Timeline reconstruction

Build a timeline from the following materials: [PASTE NOTES/DOCUMENTS]. Order every dated event, flag conflicts between sources, mark items whose dates are inferred rather than stated, and list the gaps.

82. The gap hunter

Here is a timeline of events: [TIMELINE]. What's missing? Identify unexplained gaps, periods where something must have happened but isn't recorded, and the questions each gap raises.

83. Parallel-track timeline

Lay out these related threads as parallel timelines so I can see what was happening simultaneously: [DESCRIBE THREADS]. Highlight moments where events in one thread likely influenced another.

84. Sequence-of-events interview

Help me reconstruct what happened with [EVENT]. Interview me: ask one question at a time about what I remember, starting with anchors I'm sure of (dates, documents, receipts), and build the timeline as we go.

85. Before/after comparison

Compare the state of [SUBJECT] before and after [EVENT/DATE]: what specifically changed, what stayed the same, and what changed for reasons unrelated to the event?

86. Deadline back-planner

I need [OUTCOME] done by [DATE]. Work backward from the deadline: every step, its realistic duration, dependencies between steps, and the latest safe start date for each. Flag the steps with no slack.

87. The paper-trail organizer

I have these documents related to [MATTER]: [LIST]. Organize them chronologically, tell me what the sequence reveals, and identify what documents are missing from the trail that likely exist.

88. Version-history untangler

I have multiple versions of [DOCUMENT/STORY]: [PASTE/DESCRIBE]. Track what changed between versions, when, and what each change suggests about why it was made.

89. The anniversary audit

Given this history ([SUMMARY]), what recurring dates matter going forward — renewals, deadlines, expirations, review dates? Build me a future calendar of them.

90. Narrative-vs-record check

Here's the story as told: [NARRATIVE]. Here are the dated records: [RECORDS]. Where does the narrative conflict with the record? Where does the narrative make claims no record supports?

Category 10: Fact-Checking & Verification (Patterns 91–100)

91. The claim decomposer

Break this statement into its individual checkable claims: [STATEMENT]. For each claim: is it factual, opinion, or prediction? What would verify it? How confident are you in it?

92. Full fact-check protocol

Fact-check the following: [CLAIM/TEXT]. For each claim: your assessment (supported / unsupported / false / can't verify), the reasoning, and your confidence. Do NOT assess a claim as true just because it's commonly repeated.

93. The source interrogation

I found this in [SOURCE]: [CLAIM]. Evaluate the source type: what are its incentives, typical reliability, and known biases? What kind of source would be MORE authoritative for this claim?

94. Viral-content screen

I saw this online: [PASTE]. Analyze it as a misinformation researcher would: emotional manipulation signs, missing context, misleading framing, what the image/quote/statistic would look like in full context, and how to verify it.

95. The quote authenticator

Is this quote real: "[QUOTE]" attributed to [PERSON]? Famous quotes are frequently misattributed. State whether it's verified, disputed, misattributed, or unverifiable — and don't confirm it just because it sounds like them.

96. Statistics autopsy

Examine this statistic: [STATISTIC]. Who measured it, how, on what population, and when? What definition choices could inflate or deflate it? What would the honest version of this statistic say?

97. The independent-confirmation rule

For the claim [CLAIM]: what would count as two genuinely INDEPENDENT confirmations (not sources copying each other)? Where would I find them?

98. Self-audit request (use after any important answer)

Review your last answer as a hostile fact-checker. Identify every claim that is unsupported, oversimplified, possibly outdated, or possibly hallucinated. Rate each: solid / shaky / should be verified.

99. The two-AI referee

I asked another AI the same question and got this answer: [PASTE OTHER ANSWER]. Compare it to yours: where do you agree, where do you conflict, and for each conflict, which answer is better supported and why?

100. Ground-truth finder

For settling [QUESTION] definitively, what is the authoritative primary source — the actual law, filing, dataset, official record, or document? Where is it published, and what should I search for to find it?

---

Chapter 3 explained why AI invents things. This chapter is the complete defense: a layered system you apply in proportion to the stakes.

Layer 1 — Prevention (built into the prompt)

Add guardrails before the answer exists:

Answer the question below. Rules: (1) If you are not certain, say so — never guess. (2) Separate facts from assumptions and inference. (3) Flag anything that may have changed since your training data. (4) Do not invent sources, citations, quotes, or statistics.

Question: [YOUR QUESTION]

For document work, add: "Base your answer ONLY on the document provided."

Layer 2 — Interrogation (after the answer)

The AI is surprisingly good at auditing itself when asked. After any answer that matters, use one or more:

• "Identify every unsupported claim in your answer."
• "State your confidence in each conclusion: high, medium, low — and why."
• "What evidence would change this answer?"
• "Which parts of that answer would an expert dispute?"
• "If a fact-checker reviewed that answer, what would they flag first?"

Layer 3 — Cross-examination (independent checks)

1. Two-AI check. Ask Claude and ChatGPT the same question in fresh chats. Agreement raises confidence; disagreement pinpoints what to verify. Use Pattern 99 to have one referee the other.
2. Fresh-chat re-ask. Ask the same AI in a brand-new conversation, worded differently. Hallucinations are often unstable — the story changes; real knowledge tends to stay consistent.
3. The reverse question. If it said "X causes Y," ask in a new chat, "What causes Y?" and see if X appears. Inconsistent answers = weak knowledge.

Layer 4 — External verification (leave the AI)

For anything involving money, health, law, or reputation, confirm in the real world: official websites (.gov, .org of the actual institution), primary documents, or a phone call to the organization itself. Use Pattern 100 to make the AI point you at the authoritative source.

Matching effort to stakes

• Casual curiosity (recipe idea, word origin): Layer 1 is plenty.
• Practical decisions (which product, how a process works): Layers 1–2, spot-check surprises.
• Money/health/legal: all four layers. The AI is your researcher and explainer — never your final authority.

The five red flags, revisited as reflexes

Memorize these until they're automatic: suspicious precision, uniform confidence, too-perfect examples, vague sourcing ("studies show"), and answers that flatter your question's assumptions. Any one of them = escalate a layer.

---

The source hierarchy

Not all sources are equal. From strongest to weakest:

1. Primary sources — the actual thing: the law's text, the court filing, the study itself, the official dataset, the original document, the recording.
2. Official secondary — government summaries, agency guidance, official statistics portals.
3. Expert secondary — peer-reviewed reviews, textbooks, professional bodies' guidance.
4. Quality journalism — outlets with corrections policies and named sources.
5. Everything else — blogs, social media, forwarded content, AI output itself. Useful for leads, never for conclusions.

An AI's unverified answer sits at level 5. The point of this chapter is using the AI to climb the ladder.

Prompts that climb the hierarchy

Ask for the primary source:

For [CLAIM/TOPIC], what is the primary source — the original document, law, dataset, or study? Give me its exact name and where it's officially published, so I can find it myself.

Ask for official records:

What official or government records exist about [TOPIC/ENTITY]? Which agency holds them, are they public, and how does a member of the public access them?

Ask for the scientific picture:

What does the peer-reviewed literature say about [QUESTION]? Distinguish: consensus findings, active debates, and single-study results that haven't been replicated. Name the major studies or reviews so I can look them up.

Demand independence:

News reports often copy each other. For [EVENT/CLAIM], what would genuinely independent confirmation look like — different original evidence, not the same wire story repeated?

Verifying what the AI hands you

Critical rule: AI-provided citations are themselves claims to verify. Models can fabricate realistic-looking study titles, case names, and URLs. Before relying on any citation:

1. Search the exact title — does it exist?
2. Does the real source actually say what the AI said it says?
3. Is it the current version (laws and guidelines change)?

If a web-connected AI cites a live link, open the link and check. If a citation can't be found anywhere, assume it was hallucinated.

Evaluating any source: the C-R-A-P test

• Currency — when was it published or updated? Does the topic move fast?
• Reliability — is there evidence behind claims? Corrections policy? Named authors?
• Authority — who wrote it and what's their expertise and access?
• Purpose — who benefits if you believe it? Selling, persuading, or informing?

Run a Currency-Reliability-Authority-Purpose evaluation on this source: [SOURCE/PASTE]. Be specific about incentives: who profits if readers believe this?

---

Why personas work

"Act as a [ROLE]" changes which patterns the model draws on — vocabulary, priorities, and what it checks first. A retirement question answered "as a fiduciary financial planner" surfaces different considerations than the same question answered generically.

The honest limits

A persona changes style and emphasis, not credentials. "Act as a physician" does not give the AI a medical license or your medical history. Personas make answers more useful; they don't make them more true. All Chapter 7 defenses still apply — arguably more, because persona answers sound more authoritative.

The persona library

Research Analyst

Act as a senior research analyst known for intellectual honesty. Your habits: separate evidence from interpretation, quantify uncertainty, and say "the data doesn't tell us" when it doesn't. Topic: [TOPIC]

Investigator

Act as an experienced investigator. Your habits: establish a timeline, follow the documents, note what's missing, distinguish what can be proven from what is merely suspected. My matter: [DESCRIBE]

Auditor

Act as a skeptical auditor reviewing the following: [MATERIAL]. Look for inconsistencies, round numbers that should be exact, missing documentation, and claims that don't reconcile with each other.

Scientist

Act as a careful scientist. Habits: distinguish correlation from causation, ask about sample sizes and controls, note effect sizes not just "significance," and treat single studies as provisional. Question: [QUESTION]

Physician-educator (for understanding, never for diagnosis)

Act as a physician who excels at patient education. Explain [CONDITION/TREATMENT/TEST] the way you'd explain it to a patient: what it is, what the numbers mean, what questions to ask my actual doctor. Include what you CANNOT assess without examining me.

Fiduciary financial educator (education, not advice)

Act as a fee-only fiduciary financial educator with no products to sell. Explain [FINANCIAL TOPIC] including: how salespeople typically pitch it, the incentives involved, and the questions that expose whether it fits my situation: [SITUATION]

Consumer protection advocate

Act as a consumer protection advocate reviewing this offer/contract/pitch: [PASTE]. What's designed to benefit them at my expense? What are the pressure tactics? What would you make them put in writing?

Teacher

Act as a gifted teacher who checks understanding before advancing. Teach me [TOPIC], one concept at a time, asking me a question after each to confirm I've got it.

Devil's advocate

Act as a rigorous devil's advocate. I believe: [YOUR VIEW/PLAN]. Make the strongest honest case against it — not strawmen. Then rate how much your case should worry me.

The panel (combine personas)

Convene three experts on [QUESTION]: [PERSONA 1], [PERSONA 2], [PERSONA 3]. Each gives their view, then they respond to each other's points, then you summarize where they agree, disagree, and why.

Customizing personas

Add any of these lines to sharpen a persona:

• "You have 30 years of experience and have seen every common mistake."
• "You are talking to an intelligent novice — no jargon without definitions."
• "You are candid, not polite — if something is a bad idea, say so plainly."
• "Your reputation depends on never overstating certainty."

---

Frameworks tell the AI how to organize its thinking, which reliably beats an unstructured "what do you think?" Each template below is complete — paste, fill brackets, go.

SWOT Analysis

Run a SWOT analysis on [DECISION/PLAN/SITUATION]. Context: [YOUR SITUATION]. Give Strengths, Weaknesses, Opportunities, Threats — minimum 4 each, each one specific to my context, not generic. Then: the single most important cell in the grid and why.

Five Whys

Apply Five Whys to: [PROBLEM]. Show the chain. If the chain branches, follow both branches. Stop when you reach something actionable, and distinguish root cause from contributing factors.

Weighted Decision Matrix

Build a weighted decision matrix. Decision: [DECISION]. Options: [LIST]. Propose 5–7 criteria with weights based on my priorities: [PRIORITIES]. Score each option 1–10 per criterion with a one-line justification, compute totals, then tell me how sensitive the winner is to the weights.

First-Principles Analysis

Analyze [PROBLEM/TOPIC] from first principles. Strip away "how it's usually done" and conventions. What are the fundamental facts and constraints? Building up only from those, what approaches emerge — including unconventional ones?

Risk Register

Create a risk register for [PLAN/PROJECT]. Columns: Risk, Likelihood (L/M/H), Impact (L/M/H), Early warning sign, Mitigation, Contingency if it happens anyway. Sort by severity. Then flag the risk I'm most likely underestimating.

Timeline Reconstruction

Reconstruct the timeline of [MATTER] from these materials: [PASTE]. Rules: only dated or dateable events; mark inferred dates with (~); flag conflicts between sources; list gaps where records should exist but don't.

Evidence Matrix

Build an evidence matrix for the question: [QUESTION]. Rows: each hypothesis or explanation. Columns: each piece of evidence: [LIST EVIDENCE]. In each cell: supports / contradicts / neutral. Then: which hypothesis fits best, and what single new piece of evidence would be most decisive?

Cost-Benefit Analysis

Run a cost-benefit analysis of [DECISION]. Include: one-time and recurring costs (money, time, stress), tangible and intangible benefits, when the benefits arrive vs when the costs hit, and the break-even point. State every estimate's uncertainty.

Pre-Mortem

Run a pre-mortem on [PLAN]. It's one year later and the plan failed. Write the three most plausible failure narratives. For each: earliest warning sign, and the cheapest step I could take NOW to prevent it.

The Eisenhower Sort (urgent/important)

Sort this list into the four Eisenhower quadrants (urgent+important, important not urgent, urgent not important, neither): [LIST]. Justify borderline placements. What am I treating as urgent that actually isn't?

Combining frameworks — the full workup

For big decisions, chain them in one conversation:

1. Decision matrix to structure the options →
2. Risk register on the leading option →
3. Pre-mortem on the leading option →
4. Devil's advocate persona against the conclusion →
5. Decision journal entry (Pattern 60) to record it.

Thirty minutes of this is more rigor than most major purchases and decisions ever receive.

---

Each example shows the same question three ways: Weak (what most people type), Improved (basic craft applied), Expert (full technique). Notice the pattern: each level adds context, honesty guards, and structure.

Example 1: Health

Weak:

"Is the new blood pressure medication my doctor prescribed safe?"

Improved:

"My doctor prescribed [MEDICATION] for high blood pressure. Explain in plain English how it works, common and serious side effects, and what to watch for in the first weeks. If you're unsure about anything, say so."

Expert:

"Act as a physician who excels at patient education. My doctor prescribed [MEDICATION] for high blood pressure. I'm [AGE], also taking [OTHER MEDICATIONS]. Explain: (1) how it works, in plain English; (2) common vs rare-but-serious side effects; (3) known interactions with my other medications — flagging which ones I must confirm with my pharmacist; (4) the specific questions to ask my doctor at my next visit. Separate well-established facts from things that vary by patient. Do not guess at anything — mark uncertain items 'ask your pharmacist/doctor.' This is for my understanding, not a substitute for my doctor."

Why it's better: persona, personal context, structured output, interaction check routed to the real authority, and an honesty guard.

Example 2: Finance

Weak:

"Are annuities a good investment?"

Improved:

"Explain fixed vs variable annuities for a retiree: benefits, risks, fees, and who they suit. Plain English, and note where salespeople tend to oversell."

Expert:

"Act as a fee-only fiduciary financial educator with nothing to sell me. I'm [AGE], retired, with [GENERAL SITUATION — e.g., pension plus savings], and my priority is [PRIORITY — e.g., not outliving my money]. A salesperson is pitching me a [PRODUCT]. Walk me through: (1) how this product actually makes money for the company and the salesperson — commissions and fees; (2) the realistic best and worst case for someone like me; (3) what simpler alternatives accomplish the same goal; (4) the exact questions that expose whether this fits me, and the answers that should make me walk away. Separate facts from opinion, state your confidence, and flag anything that depends on current rates, which I'll verify. This is education, not financial advice — I'll decide with a licensed advisor."

Example 3: Legal

Weak:

"Can my landlord do that?"

Improved:

"My landlord in [STATE] did [ACTION]. In general terms, what rules typically govern this, and what should I check in my lease and state law? Note that rules vary by state and you may be out of date."

Expert:

"Act as a consumer-rights educator (not my lawyer — I understand this isn't legal advice). Situation: I rent in [STATE/CITY]; my landlord [SPECIFIC ACTIONS + DATES]. Help me get oriented: (1) what area of law governs this and the general principles; (2) what specific things to look for in my lease — quote the kinds of clauses by name; (3) what my state's law typically covers on this issue, flagged as 'verify current statute'; (4) what documentation I should be gathering right now, and how to organize a timeline of events; (5) what kind of professional or agency handles this (tenant's union, legal aid, state attorney general?) and what to bring to them. Do not guess at current statutes or deadlines — instead tell me exactly where to find the authoritative version."

Why it's better: routes to primary sources and real professionals, builds the evidence file, and refuses guessed deadlines — the most dangerous kind of legal hallucination.

Example 4: News & Fact-Checking

Weak:

"Is this true? [pastes viral post]"

Improved:

"Fact-check this post: [PASTE]. Break it into individual claims and assess each: supported, false, misleading, or can't verify. Note emotional manipulation techniques."

Expert:

"Act as a misinformation researcher. Analyze this: [PASTE POST]. Steps: (1) decompose it into individual checkable claims; (2) for each: supported / unsupported / false / can't verify — with your confidence and reasoning; (3) identify framing tricks: missing context, real image with wrong caption, true statistic used misleadingly, appeals to outrage; (4) what would the full context likely show?; (5) tell me the two most authoritative places to verify the central claim, and what search terms to use. Important: do not judge a claim true merely because it is widely repeated, and say plainly if this event is after your knowledge cutoff — in which case I'll rely on the verification steps, not your assessment."

Example 5: Everyday Life (major purchase)

Weak:

"What's the best refrigerator?"

Improved:

"Recommend refrigerator types for a retired couple: reliability, cost, features worth paying for vs gimmicks. Note that specific model advice may be outdated."

Expert:

"Act as an appliance repair veteran who's seen every brand fail. I'm replacing a refrigerator: budget [RANGE], space [DIMENSIONS], two adults, priority is reliability over features. Tell me: (1) which TYPE and configuration is most reliable and why — from a repair perspective; (2) which heavily-advertised features cause the most repairs and are worth avoiding; (3) the total-cost picture: purchase, energy, likely repairs, lifespan; (4) what to check or ask before buying that most buyers don't; (5) since specific current models are beyond your knowledge, give me the checklist and search strategy to evaluate current models myself, including where reliability data actually comes from (owner surveys vs paid reviews)."

The pattern behind all five

1. Persona matched to the domain — ideally one with an incentive to be skeptical.
2. Your specific context — the answer changes with it.
3. Numbered structure — forces completeness, easy to scan.
4. Honesty guards — "don't guess," "flag what varies," "state confidence."
5. A bridge to the real world — what to verify, whom to ask, where the primary source lives.

---

Before you ask (10 seconds)

• [ ] One clear question? (Split compound questions.)
• [ ] Did I give my context — who I am, my goal, my constraints?
• [ ] Did I ask for structure (table, steps, bullets) and plain English?
• [ ] Did I add an honesty guard: "If you're not certain, say so — don't guess"?
• [ ] Is this recent-events territory? (If so: use web search or expect stale data.)

After the answer (30 seconds)

• [ ] Any of the five red flags? (Suspicious precision, uniform confidence, too-perfect examples, vague sourcing, agreeing with my slant.)
• [ ] Ask: "Identify any unsupported claims and rate your confidence in each conclusion."
• [ ] Ask: "What am I not asking that I should?"

Before you rely on it (for anything that matters)

• [ ] Cross-checked in a second AI or a fresh chat?
• [ ] Citations actually exist and say what was claimed?
• [ ] Primary source located for the load-bearing fact?
• [ ] Money/health/legal? → A qualified human is the final step, always.

The one-line version

Give context, demand honesty, check the red flags, verify what matters.

---

AI (Artificial Intelligence) — Computer systems doing tasks that normally need human intelligence: understanding language, recognizing images, making decisions.

Agent / Agentic AI — AI that can take multi-step actions (search, open files, use tools) rather than only answering in text.

Alignment — The effort to make AI systems behave according to human intentions and values.

Bias — Systematic slant in AI answers, inherited from patterns in training data.

Chatbot / Assistant — The conversational product (Claude, ChatGPT) built on top of a language model.

Claude — The AI assistant made by Anthropic. What you're likely using when you follow this guide.

ChatGPT — The AI assistant made by OpenAI.

Context window — The AI's working memory: how much of the current conversation (and uploaded documents) it can consider at once. Exceed it and early material effectively falls out of view.

Custom instructions — Standing preferences you can save (in both Claude and ChatGPT) so every chat starts with your context — e.g., "Explain things in plain English; I'm a novice."

Deep learning — The technique of training many-layered neural networks on large data; the engine under modern AI.

Fine-tuning — Additional training that specializes a general model for particular tasks or behavior.

GPT — "Generative Pre-trained Transformer," the model family behind ChatGPT. Generative = produces text; pre-trained = learned from a vast corpus first; Transformer = the underlying architecture.

Generative AI — AI that creates content (text, images, audio, video) rather than just classifying or ranking things.

Hallucination — When AI states false information as fact: invented citations, wrong dates, fabricated quotes. The central risk this guide defends against.

Inference — The act of the model producing an answer (as opposed to training, when it learned).

Jailbreak — Attempts to trick an AI into ignoring its safety rules.

Knowledge cutoff — The date after which a model's training data ends. Events after it are invisible unless the AI searches the web.

LLM (Large Language Model) — The kind of AI behind Claude and ChatGPT: a model trained on massive text to predict language.

Machine learning — Software that learns patterns from data instead of following hand-written rules.

Memory (feature) — A product feature letting the assistant retain facts about you across separate conversations. Distinct from the context window; can be turned off.

Model — A specific trained AI system (e.g., a particular version of Claude). "Newer model" roughly means "smarter and more current."

Multimodal — Able to work with more than text: images, audio, documents.

Neural network — The layered mathematical structure, loosely inspired by brain neurons, that learns patterns during training.

Open-source model — A model whose weights are published for anyone to run and modify, vs proprietary models accessed through a company's service.

Parameters — The internal numerical dials (billions of them) set during training; a rough index of model capacity.

Prompt — Whatever you type to the AI. The craft of writing good ones is prompt engineering — the subject of this guide.

Prompt injection — A trick where malicious text hidden in a document or webpage tries to hijack the AI reading it. A reason to be thoughtful about what you ask AI to process.

RAG (Retrieval-Augmented Generation) — A design where the AI first retrieves real documents and then answers from them; reduces hallucination.

Reasoning model / Extended thinking — Modes where the model works through a problem step-by-step before answering; slower but better on hard problems.

Reinforcement Learning from Human Feedback (RLHF) — Training method where human ratings teach the model which responses are good. Explains helpfulness — and some overconfidence.

System prompt — Standing instructions the AI provider gives the model before your conversation starts (its role, rules, and style).

Token — The word-chunks AI reads and writes; roughly ¾ of a word each on average. Usage limits and context windows are measured in tokens.

Training — The process of learning patterns from data — done before you ever meet the model. Your chats don't retrain it in real time.

Training data — The text corpus the model learned from; the source of both its knowledge and its blind spots.

Transformer — The 2017 neural-network architecture that made modern LLMs possible.

Web search / Browsing — The assistant's ability to search the live internet, bridging the knowledge-cutoff gap. When accuracy on current events matters, make sure it's actually searching.

---

You now have what most AI users never acquire: an understanding of why these tools fail, and a system for catching it. The three habits that matter most, if you keep nothing else:

1. Give context and ask structured questions (Chapter 4).
2. Make uncertainty welcome — "if you're not certain, say so" (everywhere).
3. Verify what matters — two AIs, then primary sources, then professionals (Chapters 7–8).

The tools will keep changing. These habits won't need to.

End of guide.