Why Most Prompts Give You Answers Instead of Thinking
There's a complaint people make about AI that goes something like this: "It gives me confident, generic, slightly wrong stuff." The instinct is to blame the model. The honest answer is that most prompts ask for output, and output is what the model returns. If you ask for a summary, you get a summary. If you ask for a list of five tips, you get a list of five tips. The result feels shallow because the request was shallow.
Researchers at Princeton studied this exact gap in their KDD 2024 paper on prompt-structure citation behavior (Aggarwal et al.). What they found is that prompt structure changes what the model attends to, not just what it says. A small shift in framing changes which sources get cited, which counterarguments surface, and which assumptions get made explicit. In other words, the prompt isn't a search query. It's a frame.
This article is the catalog of frames. Eight specific, named, copy-paste prompt patterns for specific thinking jobs. Two related Glasp pieces sit next to it. Context engineering is about assembling the room your AI works in: the documents, the sources, the constraints. The AI thinking trap is about maintaining your own critical attitude when AI starts feeling too smooth. This piece is the move catalog itself. The room, the attitude, the moves.
Each pattern below has the same structure. When to use it, where it comes from, the exact prompt, an example of what comes back, and why it works. Treat them as a vocabulary, not a checklist.
Pattern 1: Steel-Man
Steel-Man is the discipline of constructing the strongest possible version of an opposing view before you argue against it. It's the opposite of the straw man: instead of attacking a weak caricature, you defeat the toughest version.
When to use: before defending a position, publishing an argument, or making a high-stakes decision where you've already mostly made up your mind.
Origin: usually traced to Anatol Rapoport's rules of constructive debate from the 1960s, which require restating an opponent's position so well they say "thanks, I wish I'd put it that way."
Prompt:
Steel-man the opposing view of [my position]. Make the strongest possible
case, including any evidence or framing I'm likely missing. Don't hedge.
Don't add a "but actually" at the end. Just make the case.
Example output: For a position like "we should ship the MVP this week," a steel-manned response might say: "Shipping this week trades a fixable launch for a permanent first impression. Customers who bounce in week one rarely come back. The cost of a two-week delay is small. The cost of relaunching against a damaged reputation is large."
Why it works: most people don't lose arguments because their opponents are smart. They lose because they prepared against the dumb version. Steel-Man forces you to face the smart version before you're publicly committed to your side.
Pattern 2: Pre-Mortem
Pre-Mortem flips a post-mortem in time. Instead of analyzing why a project failed, you imagine it has already failed and reconstruct the path that got there.
When to use: before launching a plan, project, or feature, especially when the team already feels confident.
Origin: Gary Klein, "Performing a Project Premortem," Harvard Business Review, 2007. Klein found that teams generated 30% more reasons for potential failure when they assumed the failure was already real.
Prompt:
It's [date 6 months from now]. The plan I'm about to describe failed
badly. Walk me through the most likely sequence of events that led
to failure. Be concrete about what we missed, who pushed back too late,
and what early signal we ignored.
Plan: [paste plan here]
Example output: "By month two, the launch traffic had dropped 60% week over week. The team had assumed organic growth from referrals, but the referral loop required a feature that shipped late. By month four, the original champion at the partner company had left, and no one had built a relationship with their replacement."
Why it works: prospective hindsight changes how the brain searches. Asking "what could go wrong?" gets vague answers. Asking "what did go wrong?" gets specific ones. The fictional certainty unlocks real specificity.
Pattern 3: Load-Bearing Assumption
Load-Bearing Assumption surfaces the silent claims your argument depends on. Most arguments have one or two assumptions that, if false, collapse the whole structure. Most people never name them.
When to use: when stuck on a decision, before committing budget or headcount, or when a plan feels obviously right and you can't say why.
Prompt:
Identify the 3 load-bearing assumptions in this argument. For each, tell me:
1. What evidence would falsify it
2. How I could check in under a day
3. What changes if it turns out to be false
Argument: [paste argument]
Example output: "Assumption 1: that users will tolerate a 10-second load time on first visit. Falsifier: bounce rate over 60% on the landing page. One-day check: pull last week's analytics for the prototype URL. If false: the entire onboarding flow needs a fast-path."
Why it works: assumptions are dangerous in proportion to how invisible they are. Naming them out loud breaks the spell. The "check in under a day" constraint is what keeps the exercise honest. If you can't check it, you can't claim it.
Pattern 4: Inversion
Inversion asks the question backwards. Instead of "how do I succeed at X?", you ask "how would I guarantee the worst possible version of X?" Then you avoid those things.
When to use: when a problem feels stuck, when advice on the topic is full of platitudes, or when you can list things to avoid faster than things to do.
Origin: Charlie Munger borrowed it from the 19th century mathematician Carl Jacobi, whose advice was "Invert, always invert." Munger called it one of the most useful mental models he had.
Prompt:
Instead of solving [problem], list every action that would guarantee
the worst version of this outcome. Be specific and concrete. Then
we'll work backwards from your list to figure out what to actually do.
Example output: For "how do I write a great onboarding email?", inversion produces: "Write 800 words. Lead with company history. Bury the action button below the fold. Use 'we are excited' twice. Send on Friday at 5pm. Personalize only the first name and forget to test it."
Why it works: people are better at recognizing failure than designing success. Inversion gives the brain a concrete target it's actually good at hitting. Once the worst version is on the page, the better version is mostly the negation.
Pattern 5: Devil's Audit
Devil's Audit is the move you run on your own draft before anyone else sees it. You ask the model to play hostile reviewer, and you ask it to be specific.
When to use: after writing a draft, finishing an argument, or assembling a deck. Before sending, publishing, or presenting.
Prompt:
Audit this draft as a hostile but smart reviewer. List specifically:
1. The weakest claim
2. The most unfair generalization
3. The part where I'm assuming the reader agrees with me
4. The part most likely to be cut by an editor
Be specific. Quote the exact sentences.
Draft: [paste draft]
Example output: "Weakest claim: 'most teams fail because of culture' (paragraph 3) lacks a citation and is contradicted by the McKinsey 2023 study you cite later. Most unfair generalization: 'engineers don't read documentation.' Assumed agreement: paragraph 5 treats async work as obviously better; many readers will not."
Why it works: the gap between what you wrote and what a stranger reads is enormous. Most writers can't see their own assumptions because the assumptions feel like the air they breathe. A hostile reviewer makes the air visible. The "quote the exact sentences" constraint forces specificity, which is where the value lives.
Pattern 6: Pre-Reading Brief
Pre-Reading Brief prepares your mind before you consume a piece of content. Instead of reading passively, you read as someone who already knows what to look for.
When to use: before a long article, dense paper, important video, or book chapter you actually need to retain.
Prompt:
Before I read this, give me:
1. Three questions I should hold in mind while reading
2. The strongest counterargument I should look out for
3. The three sentences I should remember if I forget the rest
Source: [paste or link]
Example output: "Hold these questions: (1) What's the author's strongest piece of evidence? (2) Where does the timeline of events get fuzzy? (3) Whose perspective is missing? Counterargument to watch for: the author treats correlation as causation in the second study cited. Three sentences to remember: [...]"
Why it works: this is cognitive priming. Mayer's Cognitive Theory of Multimedia Learning shows that learners who get a structural preview retain roughly 30% more of what they read or watch, because they can attach incoming information to a scaffold instead of holding it in working memory cold. The brief is the scaffold.
Pattern 7: Synthesis-from-N
Synthesis-from-N is the pattern for when you've collected several sources on a topic and need to extract structure across them. Not a summary of each. A synthesis that surfaces consensus, conflict, and shared blind spots.
When to use: after reading multiple articles, papers, or transcripts on a topic. Especially useful when you have a Glasp library of highlights you've been collecting.
Prompt:
Synthesize these N sources into:
1. The core consensus (what they all agree on)
2. The loudest disagreement (where they explicitly contradict each other)
3. The assumption all sources share that nobody questions
Sources: [paste highlights, links, or quotes]
Example output: "Core consensus: all five sources agree that working memory is the bottleneck for learning. Loudest disagreement: sources 2 and 4 split on whether spaced repetition outperforms interleaving. Shared unquestioned assumption: every source treats motivation as exogenous and ignores how the format itself shapes engagement."
Why it works: most people read N sources and end up with N summaries. Synthesis-from-N forces a single map. The third bullet, the unquestioned assumption, is where the actual insight usually lives, because that's the question nobody in the field is asking yet. For a longer treatment of the read-highlight-synthesize cycle, see the synthesis loop.
Pattern 8: Failure-Mode Hunter
Failure-Mode Hunter is the systems-thinking pattern. You list the ways the thing can break, rank them by probability and severity, and ask what would catch each one early.
When to use: when designing systems, processes, products, or tools where reliability matters more than novelty.
Prompt:
List the top 7 failure modes for [system]. For each, give me:
1. Probability (low / medium / high)
2. Severity (low / medium / high)
3. The cheapest detection mechanism that would catch it within an hour
System: [describe system]
Example output: "Failure mode 1: database connection pool exhausted under traffic spike. Probability: medium. Severity: high. Detection: alert on connection wait time over 200ms. Failure mode 2: third-party API silently rate-limits without erroring. Probability: high. Severity: medium. Detection: synthetic check that compares response payload to a known fixture every 5 minutes."
Why it works: failure modes are usually known to someone. The question is whether they're known to you, before they happen. Forcing probability, severity, and a cheap detection mechanism turns vague worry into a prioritized checklist. The "within an hour" constraint is what separates real monitoring from theater.
A One-Page Cheat Sheet (and How to Memorize the Patterns)
Here's the whole catalog in one place.
| Pattern | When | Trigger Phrase |
|---|---|---|
| Steel-Man | Before defending a view | "Make the strongest case against me." |
| Pre-Mortem | Before launching a plan | "It already failed. What happened?" |
| Load-Bearing Assumption | When stuck or committing | "What 3 assumptions hold this up?" |
| Inversion | When success feels vague | "How would I guarantee the worst?" |
| Devil's Audit | After writing a draft | "Be the hostile reviewer." |
| Pre-Reading Brief | Before consuming content | "What 3 questions should I hold?" |
| Synthesis-from-N | After reading several sources | "Consensus, conflict, shared blind spot." |
| Failure-Mode Hunter | When designing a system | "Top 7 failure modes, ranked." |
The trick to memorizing them is to name them in your own work. The next time you ask for the strongest counterargument to your position, don't just type it. Say "I'm running a Steel-Man on this." When you finish a draft, say "running the Devil's Audit." Naming the move turns it from a one-off prompt into a tool you can pick up consistently.
A practical rhythm that works well with Glasp's web highlighter: highlight first, pattern second. As you read, save the passages that feel important. Then run patterns over your highlight library: Steel-Man your position using highlights from sources you disagree with, Devil's Audit a draft against highlights you've gathered, and Synthesize-from-N across a topic cluster. The highlights are the raw material. The patterns are the moves you make on the material. Glasp's AI chat feature is built around this rhythm. Your highlights are already in context, so the patterns operate on what you've actually read, not what the model guesses.
Most people who get good at this never use all eight in a single week. They pick two or three that fit how they think, and they get fluent. That's the goal. Not eight patterns memorized. Two or three patterns running on autopilot.
Frequently Asked Questions
Do these patterns work in any LLM?
Yes. They've been tested across GPT-5, Claude 4.6, and Gemini 2.5, and they hold up because they're about prompt structure, not model-specific tricks. The example outputs above were generated across all three with similar quality. Smaller models will produce shallower versions of the same shape, but the shape carries.
Should I memorize all 8?
No. Start with two. The two with the highest leverage for most knowledge work are Steel-Man and Pre-Mortem. Steel-Man saves you from defending positions you haven't actually pressure-tested. Pre-Mortem saves you from launching plans whose failure mode is already obvious in retrospect. Add the others as the situations come up. Inversion is the third one most people end up adopting.
Is this just prompt engineering?
Related, but the goal is different. Prompt engineering optimizes for the model's output: better answers, fewer hallucinations, cleaner formatting. Thinking patterns optimize for your cognition: catching your own blind spots, surfacing your own assumptions, sharpening your own arguments. The output is a side effect. The point is what happens in your head when you run the pattern.
What about chain-of-thought or step-by-step prompts?
Chain-of-thought is a meta-pattern that can stack with these. You can append "think step by step" to any of the eight patterns above and usually get a more rigorous response, especially in models that don't already do reasoning by default. But CoT alone tends to produce verbose thinking aimed at no particular target. The eight patterns above point the thinking at something specific. Use them together when the stakes warrant the extra tokens.
Conclusion
The complaint that AI is "shallow" is mostly a complaint about shallow prompts. The patterns above don't make the model smarter. They make your request smarter, which is where the leverage actually lives.
Eight named moves. Steel-Man for opposing views. Pre-Mortem for plans. Load-Bearing Assumption for decisions. Inversion when stuck. Devil's Audit on drafts. Pre-Reading Brief on content. Synthesis-from-N across sources. Failure-Mode Hunter on systems. The names matter because they're what you'll actually remember in the moment. "Run a Steel-Man on this" is a thing you can say to yourself. "Use a more critical framing in your prompt" is not.
If you take one thing from this catalog, take this: the goal is not better answers from AI. The goal is sharper thinking from you, with AI as the partner that asks the questions you forgot to ask yourself. The eight patterns are eight reliable ways to make that partnership work. Pick two. Get fluent. Add more when the work demands it.