Direction and discernment — AI Literacy Framework

What shapes the AI’s answer

You control most of these inputs. Good direction makes good outputs.

Three things follow from this. Memory is why the privacy check in Part 2 mattered. Earlier messages steer later answers, which is why Test 2 below needs a fresh chat. And files you upload are inputs, so whatever is in them goes to the tool.

Good direction has three parts

Product

What you want: the output, context, audience, format and length.

“A half-page flyer for Grade 10 students, plain language, three short sections.”

Process

How to get there: steps, sources to use, examples, what to leave out.

“Use only the notes I’ve pasted. Ask me three questions before you start.”

Performance

How the AI should behave: role, tone, pushback.

“Act as a skeptical youth worker. Tell me what won’t work.”

Performance is the new one for most people: telling the AI how to behave.

You don’t actually need to be any good at prompting to be good at directing AI

For example, ask it to help:

Help me create a good prompt to do <X>. Ask me up to five questions, one at a time, to help shape this prompt.

The direction–discernment loop

The loop: direction and discernment, repeated until the output meets your needs.

Direction — steer it

Product: what you want: context, task, audience, format
Process: how to get there: steps, examples, what to leave out
Performance: how the AI should behave: role, tone, pushback

Discernment — judge it

Hallucination: is it true? Confident answer ≠ correct answer
Bias: whose view is built in, and whose is missing?
Taste: is the output actually good? Does it match what you want?
Flattery: is it right, or just agreeing with me?

Direct, judge, redirect, until it meets your needs and expectations.

The four tests below each put one of these to work.

Test 1 — Ask about something you know very well

Test 2 — Many stories, similar pattern

Open a fresh chat. Type this exactly:

Write a one-paragraph story about a surgeon and a nurse. Make sure to give them names

Then tally the gender of each role, the names, the ages and the setting. In a group, put everyone’s stories side by side. On your own, run the prompt three or four times, each in a fresh chat.

Then judge: which story is best, and why? What would make the others less predictable?

What to expect

Each story looks unremarkable on its own. Side by side, a pattern usually shows: the male surgeon, the female nurse, the obvious names. The stereotyped stories are usually also the dullest, because the default is often both biased and boring.

Bias is a pattern, so you won’t necessarily see it in your own output.

Test 3 — Everyone’s original idea

Five steps

Humans first — one minute, no AI: brainstorm a campaign to recruit youth volunteers.

Generic prompt — ask the AI:

Give me a creative idea for a campaign to recruit youth volunteers for a community organization.

Redirect — rewrite the prompt with product, process and performance, using your real organization and community. Compare again.

Go deep — paste the 100-ideas prompt. It asks for 100 ideas, then the five boldest.

I need ideas for a campaign to recruit youth volunteers for a local community organization. The ideas don’t need to be proven or even clearly feasible. Follow these steps. Do each step, even if you think you don’t need to. First, generate a list of 100 ideas (short title only). Second, go through the list and decide whether the ideas are different and bold; change them as needed to make them bolder and more different. No two ideas should be the same. This is important! Next, choose the five most different ideas, give each a name, and describe each in 40–80 words. Do this step by step!

Adapted from Meincke, Mollick and Terwiesch (2024), “Prompting Diverse Ideas: Increasing AI Idea Variance.”

Choose — of everything on the table, which idea would you actually run? Say why in a few words. That’s taste.

What to expect

The generic prompt gives strikingly similar ideas to everyone who runs it, while ideas from people are more scattered and stranger. After the redirect, the AI’s ideas should start to diverge. The 100-ideas prompt gets far more variety, mostly because it describes a process: many ideas first, then a push to make them bolder and more different.

Notice your reasons for choosing, like “it fits our town” or “a 16-year-old would actually do this”. Those reasons are taste, and they’re also what made your redirected prompt better.

Generic in → generic out. Specific context and coaxing in → unique results out.

Test 4 — A thought partner who pushes back

This is the model of good use. The AI does no writing for you. It questions your reasoning so your own thinking gets stronger.

Identify, paste, back and forth

Identify — write down a challenge or problem you’ve been thinking about: school, work or life. Note what you’re leaning toward, if anything. Leave out private details about other people.

Paste — paste the thought-partner prompt and replace [describe it here] with your challenge, in the same message.

You are my critical thought partner. Below is a challenge or problem I’ve been thinking about, and what I’m leaning toward, if anything. Help me think it through by challenging my reasoning. Be blunt, concise and skeptical. Skip praise. Probe how I’ve framed the problem, my assumptions, evidence, trade-offs and missing perspectives, especially those of the people affected. Press evasive answers; change your view when the reasoning or evidence warrants it. Don’t invent flaws or argue just to disagree. Separate facts, inferences and guesses. Start with a brief critique: the weakest point in my thinking, hidden assumptions, what could go wrong, what’s missing. Then ask the most important question I need to answer, and continue one question at a time. Keep the decision with me. Don’t solve it for me or rewrite my work unless asked.

My challenge: [describe it here]

Back and forth — answer its questions one at a time. Push back when you disagree. If it gets too agreeable, add:

You have mostly agreed with me. Identify the assumption most likely to fail, explain the consequences and ask me a question that would help test it.

The AI asks the questions. You make the decisions.

Discernment summary

Each test showed one lens. Here are all four together.

HallucinationTest 1	Is it true?	Check claims, numbers, quotes and references against a real source. Take most care where you know least.
BiasTest 2	Is it fair?	Whose view is built in, and who’s missing? Run it more than once, or compare with others.
TasteTests 2 and 3	Is it actually good?	Right for the audience and purpose? Specific, not the default? Would you put your name on it?
FlatteryTest 4	Is it just agreeing?	Ask it to argue the other side. Be wary of praise.

Fluent isn’t the same as true, and agreeable isn’t the same as good.