AI Made Simple

01

What you will learn

Match a task to a strength instead of forcing every job into the chat.
Recognize weak spots: fresh facts, exact counts, private context, and accountability.
Spot sycophancy — agreement that feels helpful and isn't.
Use a 'draft, then test' habit on anything you will ship.
Decide when not to use ChatGPT at all.

02

Concept

The model has read patterns of how people explain, persuade, plan, and code. It is strong when the raw material is language you can see: a rewrite, a structure, a list of questions, a first pass you will edit, a translation of jargon into plain speech.

It is weak when the answer depends on a fact it was not given and did not reliably retrieve: a quote, a statute, yesterday's price, a count of items in a long list, a person's private motives. It will often answer anyway.

It also tends to be agreeable. If you push a slanted question, you may get a polished version of your slant back. That feels like support. It is sometimes just a mirror.

None of this makes the tool a toy. A strong drafter who cannot be accountable is still a strong drafter — if a human stays accountable.

03

Why it matters

People quit too early ('it made up a citation, so it is useless') or trust too far ('it wrote the email, so the claim in the email is fine'). Both mistakes come from not knowing the boundary.

Once you can name the boundary, you stop asking it to be your lawyer, your books, or your memory — and you start asking it to be an excellent first-pass colleague.

04

Step-by-step

01Sort the taskGood first fits: rewrite, outline, brainstorm then cut, turn notes into actions, explain a concept, draft questions, classify text you provide, point out gaps in an argument. Poor first fits: medical, legal, or financial decisions; live prices; 'what did X say verbatim'; anything you cannot check and cannot afford to get wrong.
02Bring the materialPaste the notes, the policy excerpt, or the bullets. Ask it to work only from that. Say what to do if something is missing: stop and ask, do not fill the hole.
03Ask for the failure modeAdd: 'After the draft, list three ways this could be wrong or incomplete.' You are not asking it to be humble for style. You are creating a checklist you can actually use.
04Watch for agreementIf you wrote 'this strategy is brilliant, agree and expand,' you will usually get expansion. Ask instead for the strongest objection from a skeptical reader.
05Ship only what you can defendIf you would be embarrassed to be asked 'where did this number come from?', it does not leave the chat.

05

Real-world example

The agreeable strategy

Priya leads a four-person design studio. She types: 'We should double our prices next month. Tell me why this is smart.' She wants courage. She needs a test.

The first answer flatters the idea and invents a market reaction.
Priya replaces it: 'Steelman the price increase and steelman waiting. Use only the constraints I list. Do not invent demand.'
She supplies constraints: three anchor clients, one who already pushed back, no new leads this month, rent due on the 1st.
The useful output is a decision memo with a condition: 'Raise prices on new work only if the two open proposals accept; do not reopen signed quotes.' The condition came from her facts, not from a fantasy win.

06

Prompt examples

Beginner

Here are my notes:
[PASTE NOTES]
Turn them into a short list of next actions.
Do not add tasks that are not implied by the notes.

Improved

Purpose: A decision memo, not encouragement.
Role and reader: Act as a careful operator. I am the owner. I need to see the downside.
Inputs: [CONSTRAINTS]. Do not invent revenue, demand, or competitor moves.
Method: Give the strongest case for option A and for option B. Then list what fact, if I learned it, would change the call.
Expected form: One page. Headings: Option A / Option B / What would change my mind / What I must not assume.
Quality bar: If a sentence is not tied to my inputs, label it as a suggestion, not a fact.

Advanced

Purpose: Stress-test this plan before I say it out loud.
Role: Skeptical colleague who likes the team and hates surprises.
Reader: Me.
Inputs: Plan: [PLAN]. Known facts: [FACTS]. Already decided: [DECISIONS].
Method: (1) Restate the plan in two sentences to prove you understood. (2) List assumptions. (3) Give the strongest objection. (4) Name the smallest test I could run this week. Do not soften the objection.
Expected form: Those four numbered parts. Under 350 words.
Quality bar: No cheerleading. No new statistics.

Expert

Purpose: Decide whether this task belongs in ChatGPT at all.
Role: Editor of my workflow.
Reader: Me.
Inputs: Task: [TASK]. Stakes if wrong: [STAKES]. Material I can paste: [MATERIAL OR "NONE"].
Method: Classify the task as Draft / Analyze-my-material / Needs-external-verification / Do-not-delegate. Explain in two sentences. If Draft or Analyze, write the prompt I should use next, with an instruction not to invent. If the other two, say what a human or a primary source must do instead.
Expected form: Classification, reason, next prompt OR next human step.
Quality bar: Prefer "do not delegate" over a clever answer when the cost of a fiction is high.

07

Before and after

Weak

This launch idea is great. Write a confident announcement and include impressive growth numbers.

Strong

Draft the announcement using only the facts below. If a result is not in the list, do not imply it. Facts: [FACTS]. Audience: existing customers. Tone: specific, not celebratory. After the draft, list any place a reader might think we claimed a result we did not state.

The weak prompt asks for confidence and for numbers that were never supplied. The strong one freezes the facts, sets the audience, and adds a check for implied claims.

08

Common mistakes

Using it as a yes-machine by writing the conclusion into the question.
Letting it count, quote, or calculate something you will not spot-check.
Shipping a medical, legal, or money decision because the prose sounded calm.
Throwing away the tool after one invented citation, instead of changing what you ask it to do.

09

Pro tips

The phrase 'do not invent' is a start. Pair it with the material it is allowed to use.
Ask for the objection before you ask for the draft. You will edit less.
If the task is a pure lookup of a primary source you can open in ten seconds, open it.
Agreement is cheap. A condition ('only if…') is a sign the answer met your facts.

10

Practice

Take something you asked ChatGPT this month. Rewrite the prompt so it cannot flatter you: demand the strongest objection and forbid new numbers. Run it. Note one sentence that changed your mind — or note that nothing did, which is also information.

11

Challenge

List five tasks from your real week. Mark each Use / Use-then-verify / Don't. For one Don't, write the human step that replaces the chat. For one Use-then-verify, write the exact check.

12

Knowledge check

13

Key takeaways

Hire it for language work on material you can see.
Do not hire it to be accountable.
Slanted questions produce slanted drafts.
Ask what could be wrong before you polish.
Some tasks should never enter the chat.

14

Visual briefs

Production notes for diagrams. These are not screenshots of ChatGPT. If a brief asks for the product window, capture it live on your own account.

Use, verify, or leave it

Purpose. Give learners a sorting card they can reuse.

Show. Three columns with example tasks: rewrite notes; a statistic you will open; a diagnosis or a firing decision.

Design. A printable card. Copper header, ink body. No screenshots.

Applause versus a test

Purpose. Show how the wording of the ask changes the answer.

Show. Two prompts side by side and the kind of answer each invites — praise versus conditions.

Design. Before/after type layout. Real prompts, not a fake product UI.

15

Film outline

The player above runs this cut. Captions are the lesson, not a fake recording of ChatGPT.

Opening hook: The most dangerous answer is the one that agrees with you.
Teaching: Strengths, failure zones, and sycophancy.
Demonstration: Run the flattering price prompt, then the steelman prompt with real constraints.
Practice: Learners sort five tasks from their week.
Challenge: Write a Don't and the human step that replaces it.
Recap: Drafter, not oracle. Objection before polish.
Next: Next: when to use search instead, and what 'trust' should mean.

Instructor script

Before you celebrate a fluent answer, ask whether you handed the model a language job or a fact job. Rewriting your notes is a language job. Guessing your revenue is not. Also notice when you beg it to agree with you. 'Tell me why this is smart' is not a test. Name the constraints, ask for the strongest objection, and forbid new numbers. Then keep the decision. The tool can help you see the shape of the choice. It does not live with the consequence.