LLMs as a Thinking Partner

The single best habit to build is asking the model to push back. Left alone, assistants are agreeable - they will happily polish a weak idea. Their real value as a thinking partner is unlocked when you explicitly ask them to challenge you, because a good objection you had not considered is worth more than a compliment you did not need.

The move is simple and endlessly reusable:

I am proposing a central double-height courtyard as the organising
idea for this school. Play devil's advocate: give me the 6
strongest arguments AGAINST this move - climate, cost, acoustics,
supervision, circulation, maintenance - and for each, the condition
under which it becomes a deal-breaker. Do not spare my feelings.

This is pre-mortem thinking on demand. You are not asking the model to decide - you are asking it to surface the objections a sharp critic or a demanding client would raise, so you can answer them before they do. Related plays: "steelman the opposite approach," "what am I assuming here that might be wrong?", "what would make this fail on site?", "a jury will ask me three hard questions about this scheme - what are they?" The model is a stand-in for the critical audience you do not have in the room. And because it does not tire or take sides, you can stress-test an idea ten times over without wearing out a colleague's patience - then bring your judgement to bear on which objections actually bite.

Design thinking is often a swamp of half-formed thoughts, and getting them into order is real cognitive work. An LLM is an excellent externalising surface - the modern, talking version of the 'rubber duck' programmers use, where explaining a problem out loud is what solves it. Dump the mess in and put it to work on the shape.

Useful structuring moves: "here are my scattered notes from the client meeting - organise them into themes and flag what is missing," "turn this rambling paragraph into a clear argument with a claim and three supports," "I have four competing priorities for this plan; help me build a simple matrix to weigh them," "summarise what I actually decided in this thread." The model does not supply the judgement - you decide which theme matters, which priority wins - but it takes the friction out of getting from chaos to a structure you can think against.

Thinking out loud with it is legitimate and powerful. Talk through a problem the way you would with a colleague - "I am torn between two circulation strategies; let me reason through both and you point out where my logic wobbles." The value is partly the model's responses and partly the act of articulating: writing forces clarity, and the model's questions keep you honest. Use voice input if it helps you think aloud. This is where the LLM most resembles a genuine collaborator - not because it is intelligent in the way a person is, but because it gives your own thinking something to push against, instantly and without judgement.

As a learning partner, an LLM is close to a private tutor - available always, infinitely patient, adjustable to your exact level. This is one of its most reliable, high-value uses, because you can verify what you learn as you go. Adjust the level explicitly: "explain how a chilled-beam system works to an architect who knows the basics but not the engineering," "give me an analogy for embodied vs operational carbon," "quiz me on the fire-egress principles I just read, one question at a time." Ask follow-ups without embarrassment, request analogies, ask it to relate a new concept to something you already know. The one caution from Lesson 1.1 stands: for anything factual it may be confidently wrong, so treat it as a brilliant tutor who occasionally misremembers - cross-check the load-bearing facts, especially codes and numbers.

As a drafting partner, the point is not the draft - it is having something concrete to react to. A blank page is expensive; a mediocre first draft you can attack is cheap and useful. "Draft a rough project narrative from these bullet points so I have something to rewrite," "give me three opening lines for this concept statement in different registers," "outline this report so I can fill it in." You will often find you know exactly what you think the moment you see a version you disagree with - the draft is a foil for your judgement, not a replacement for it. That reframing matters: you are not asking the model to write your work; you are using it to get past the friction of starting, then doing the real authoring yourself.

A thinking partner is only as useful as your clear sight of what it is not, and a fluent, confident model makes those limits easy to forget. Name them, so they do not quietly mislead you. First, sycophancy: assistants are trained to be agreeable, so their default is to validate you. Left unchecked this is dangerous for thinking - it will reassure you that a weak scheme is strong. The whole devil's-advocate habit exists to fight this, but you have to invoke it every time; the model will not volunteer to disagree.

Second, no lived experience and no stakes. The model has never stood on a site at 4pm watching where the sun falls, never sat with a client who cannot articulate what they hate, never carried the consequences of a decision. It can recombine what it read, which is genuinely useful, but it cannot bring judgement forged by responsibility - and its suggestions cost it nothing if they are wrong. Weigh its input accordingly: strong on breadth and articulation, empty on the tacit, embodied knowledge that separates a good designer from a well-read one.

Third, it regresses to the mean. Because it predicts what is most typical, its unaided ideas tend toward the conventional, the average, the already-common - the opposite of the genuine novelty design prizes. It is a fine partner for stress-testing and structuring a bold idea you bring; it is a poor source of the bold idea itself. Fourth, the confident-wrong problem from Lesson 1.1 does not go away just because the mode feels collaborative - a plausible-sounding argument can rest on a fabricated premise, so the facts inside its reasoning still need checking.

None of this cancels the value; it calibrates it. The right mental model is a brilliant, widely-read, endlessly patient sparring partner who is also a bit of a people-pleaser, has never actually built anything, and occasionally states falsehoods with total conviction. Enormously useful with those limits in mind - quietly misleading without them.

Everything above has one precondition, and the whole module rests on it: you stay the thinker. There is a real, well-documented risk with a fluent, always-agreeable partner - you can quietly outsource your own thinking to it. You stop forming your own view because a plausible one is always on tap; you accept its framing because it arrived first; its confident, well-written answer crowds out the harder, better idea you would have reached by struggling a little longer. For a student especially, leaning on it instead of building the underlying skill is a slow, invisible loss.

Guard against it deliberately. Form your own view first, then bring it to the model to test - do not let it think before you do. Treat its output as a proposal, not a verdict - the point of asking for the counter-argument is that you then judge which objections hold. Notice the pull of the agreeable answer and push against it: the model's job here is to make your thinking sharper, not to make your decisions. Keep the skills you will need without it - if you never struggle with the blank page or the hard analysis, you never build the muscle, and the whole edifice of judgement this course depends on gets weaker.

The test is simple: after a session with the model, are you thinking more clearly and holding a view you can defend as your own - or are you just relaying its answer? The first is a thinking partner amplifying you. The second is autopilot wearing a clever disguise. Used with that line held, the LLM is the highest-leverage everyday tool in this whole course: it makes you a faster, sharper, better-prepared designer - precisely because it never gets to be the designer.