AI as a Thinking Partner

Most knowledge work begins with "I need to understand X." LLMs are extraordinary research partners when used well — and dangerously misleading when used poorly. This lesson covers both halves.

The fundamental tension

LLMs know a lot, present it confidently, and sometimes make things up. The same property — confident generation — is both their strength (fast, articulate answers) and their failure mode (plausible nonsense). Research with AI is the discipline of getting the speed without the nonsense.

The "explain it to a smart skeptic" prompt

For getting a clean explanation of an unfamiliar topic, this is the highest-leverage move I know:

Explain {topic} to me. I'm technically literate but new to this area.
Assume I'll push back on anything that sounds like marketing or wishful
thinking. Include:

- The core idea in one sentence
- The problem it actually solves (vs. the problem people claim it solves)
- The strongest argument against it from informed critics
- Three concrete examples — at least one where it doesn't work well
- The two or three things I should learn next to go deeper

The structure forces the model to give a balanced view rather than the marketing pitch. The "informed critics" line is especially important — it pulls the model out of its default mode of helpful enthusiasm.

The compare-and-contrast prompt

For understanding the landscape of options:

I'm deciding between {A} and {B} for {use case}. Give me:

1. The summary version of each, two sentences max
2. The thing each does better than the other
3. The thing each does worse than the other
4. Which I should choose if my main constraint is X
5. Which I should choose if my main constraint is Y
6. The third option I should also consider that I didn't mention

That last item — "what should I also consider" — regularly surfaces useful options I wouldn't have known to ask about.

Synthesis across sources

When you have multiple sources on a topic — interview notes, research papers, customer feedback — AI is excellent at finding themes you'd miss reading sequentially.

I have notes from 8 customer interviews about why people churn. Read
all of them and tell me:

1. The 3-5 themes that show up most often
2. For each theme, which interviews mention it (by ID)
3. Quotes that capture each theme well
4. The themes that show up in only one or two interviews but might be
   important signals
5. What's notably absent — what surprised you NOT to find?

<interviews>
{interview_notes}
</interviews>

The "what's notably absent" question is a research multiplier. Models are good at it; humans rarely think to ask.

The trust-but-verify discipline

For any research output you're going to act on:

1. Cross-check the claims that matter. If a fact would change your decision, verify it against a primary source. Don't trust the LLM's footnotes — verify the underlying claim.
2. Ask for the strongest counter-argument. "What would someone who disagreed with this say?" Counter-balances the confident-positive default.
3. Distinguish what the model knows from what it inferred. "How confident are you in each claim? What's well-established vs. your inference?"
4. Be especially suspicious of: dates, numbers, citations, names, quotes. These are where confident hallucinations cluster.

The "kindly disagree with me" prompt

One of the most useful research prompts isn't research at all — it's structured pushback:

Here's my current thinking on {topic}: {your view}.

Steelman the opposite position. What would someone smart and well-
informed say if they disagreed with me? Don't try to convince me —
just give me the strongest version of the other view.

This catches blind spots cheaply. Use it before any decision you've already half-made.

Most people use AI for writing in the worst possible way: paste topic, copy output, post. The result is recognizable — flat, generic, voice-less. This lesson is about the techniques that actually improve your writing rather than launder it through a model.

The fundamental rule

The AI should never be the writer. You are the writer. The AI is the editor, the brainstormer, the sounding board, the tireless reviewer. The moment you let it write the thing, the thing stops being yours and starts being a draft anyone could have produced.

This rule has practical consequences. Don't paste "write a blog post about X" prompts. Write the draft yourself, however roughly, then use AI to make it better.

Brainstorm structure before content

Where AI genuinely accelerates writing: getting from "I have a vague topic" to "I have an outline I want to write." Use it for that, then go solo for the actual writing.

I want to write a piece about {topic}. My audience is {audience}. My
goal is to {goal}. My personal angle is {your perspective}.

Give me three different outlines I could use. For each, list:
- The opening hook
- The main argument arc
- The closing
- What kind of reader each outline would resonate with most

Don't write the post. Just the structures.

You pick one outline, modify it to taste, then write the actual piece in your own voice.

Use AI to find the boring parts

After you draft, AI is an excellent first pass at identifying weakness:

Read this draft. Tell me, specifically:

1. The single most boring paragraph — the one a reader is most likely
   to skip
2. The weakest claim that's stated as if it's strong
3. Anywhere the voice slips into corporate / generic phrasing
4. One missing counter-argument that would strengthen the piece if
   I addressed it

Don't rewrite. Just tell me where to look.

<draft>
{your draft}
</draft>

"Don't rewrite" is critical. The moment you ask the AI to rewrite, you get its voice instead of yours. The point is to use the AI as a critic, not a ghostwriter.

The voice-preservation move

If you do want help rewriting a passage, include examples of your own writing as the style anchor:

Rewrite the paragraph below to be sharper and tighter. Match the voice
of these examples I wrote:

<style_examples>
{three paragraphs of your own writing}
</style_examples>

<paragraph_to_rewrite>
{the bit you want help with}
</paragraph_to_rewrite>

This won't perfectly clone your voice, but it'll prevent the model from defaulting to its bland house style.

Banned words and phrases

Every model has tells. Phrases that show up over and over in AI-written text. Knowing yours and explicitly banning them is a small but powerful move.

Common LLM tells (varies by model):

• "It's worth noting that..."
• "In today's fast-paced world..."
• "Delve into / dive deep into"
• "In conclusion / In summary"
• "Robust" used generically
• Em-dash sandwiches separating every clause
• "Not just X, but Y" as a hook construction

Add a "do not use these" list to your editing prompts. It catches a meaningful fraction of the AI-sounding phrasing.

The "explain this back" check

For any non-trivial piece of writing, this is gold:

Read this and explain back to me, in your own words, what you think
my main argument is and what evidence I'm using to support it.

Don't critique. Don't suggest changes. Just tell me what you think
the piece says.

<piece>
{your writing}
</piece>

If the explanation diverges from what you meant, the writing is unclear. The model is acting as a fresh reader. This catches "I knew what I meant" blindness.

When AI is the wrong tool

AI is the wrong tool for:

• Anything where your unique experience is the value (personal essays, founder stories)
• Writing where being slightly wrong is catastrophic (legal, medical, financial advice)
• Writing where the process IS the point (journaling, working through your own thoughts)
• Anything you'd be embarrassed to admit you used AI for

The last one is the honest gut check. If you wouldn't tell a colleague how the piece was made, the piece is worse for the AI being in it.

The least-discussed and possibly highest-leverage use of AI: as a thinking-partner for decisions. Not "AI decides for me" — that's a category error. AI as a structure for your own decision-making, ensuring you don't skip steps, that you've considered the alternatives, and that you've stress-tested your assumptions.

The reframing prompt

Most decisions are pre-loaded with framing. "Should I take this job?" already presumes the choice is "take it / don't." Often the better answer is a third thing nobody framed.

I'm trying to decide between {A} and {B}. Before I evaluate those:

1. What's the broader question underneath this choice?
2. What are 2-3 other framings of this decision that I might be
   missing?
3. Is there a third option — something I haven't considered — that
   would obviously dominate both?

The point isn't to invalidate the original framing. It's to make sure you've checked.

The pre-mortem

Standard decision-making asks "will this work?" The pre-mortem asks "imagine it failed — what happened?"

I'm planning to do {thing} because {reasoning}.

Imagine it's 6 months from now and this clearly didn't work. Write
me a brief postmortem from that future. What happened? What did I
miss? What signs should I have seen earlier?

Give me three different failure scenarios. The most likely one, a
boring one, and a surprising one.

The "boring one" matters. Most things fail not because of dramatic risks but because of mundane ones — schedule slip, team turnover, second-priority items getting deprioritized. AI is good at surfacing the boring failures because it has seen so many of them.

Structured pro/con with weighting

The "list pros and cons" trick is older than computers. Adding AI helps in one specific way: enforcing structure and surfacing what you'd skim past.

I'm deciding whether to {decision}. Here's my current thinking:
{your thinking}.

Build me a decision matrix:

| Factor | Weight (1-5) | Option A score | Option B score | Notes |

Choose the factors yourself based on what's relevant. Be honest about
the weights — if a factor sounds important but I clearly don't care
about it, weight it low. Highlight the factor where the two options
differ most.

The final question is the prize. The factor where options differ most is usually the actual deciding factor — even if you'd been treating other factors as more important.

The "what would change my mind" prompt

For any conclusion you've reached:

I've concluded {conclusion} because {reasons}.

What evidence would change my mind? Specifically:
1. What's something I could learn that would make me less confident?
2. What's something that would flip me to the opposite view?
3. What's the next thing I should investigate to find out if either
   of those is true?

This is calibration training for your own beliefs. If you can't name what would change your mind, your "conclusion" is a feeling pretending to be a conclusion.

Decision rehearsal

For decisions you're going to have to defend (asking for budget, presenting a proposal, advocating internally):

I'm about to propose {thing} to {audience}. Help me rehearse.

1. Play the role of a skeptical {audience}. Ask me the 5 hardest
   questions you'd ask.
2. For each, write a brief answer in my voice.
3. Identify which of my answers are weakest — where I'd fold under
   follow-up pressure.

Cheap rehearsal of conversations you'll have anyway. Catches gaps before they cost you in the real meeting.

What AI is NOT good for in decisions

Be honest about the limits:

• Decisions that depend on tacit knowledge of your specific situation. AI knows generic. It doesn't know your team, your relationships, your gut.
• Highly personal trade-offs. AI can structure your thinking; it can't tell you what matters to you.
• Decisions where you want validation, not analysis. Sometimes you've already decided. That's fine — but don't pretend you're "asking AI for advice."
• Predictions about specific future events. The model is interpolating from training data, not seeing the future.

The compounding effect

None of these prompts are individually revolutionary. The compounding effect is. Over a year, an engineer who runs pre-mortems on every meaningful decision, reframes the question before evaluating, and rehearses defenses before advocating — that engineer makes meaningfully better decisions than one who doesn't. Not because they're smarter. Because they have a thinking partner that never gets tired.

Module 8 wrap-up

This module isn't about building AI products. It's about being a better thinker because you have AI as a partner. The techniques here are simple — almost embarrassingly simple. The discipline of actually using them is what compounds.

Course wrap-up

Eight modules, twenty-four lessons. From the foundations of how LLMs work, through the techniques and patterns that make them reliable, into production engineering, evaluation, safety, AI-assisted coding, and using AI to think.

The one habit, if you take nothing else: treat every prompt as something you can measure, iterate on, and improve. That's the difference between people who get lucky with AI and people who consistently get value from it.

Now go build something.