Architectural Vocabulary for AI

The single rule under this whole lesson: the model responds to words that name a bounded, learnable cluster, and ignores words that don't. 'Travertine' names one stone with a consistent look across millions of captions, so it lands hard. 'Beautiful', 'modern', 'stylish', 'high-quality' were attached to every kind of image imaginable, so they average out to a faint, directionless nudge.

This is why prompts full of praise underperform prompts full of nouns. When you catch yourself reaching for an adjective of taste, stop and ask: is there a named thing I mean instead? 'A beautiful facade' becomes 'a board-formed concrete facade with deep reveals'. 'A cozy room' becomes 'a room with warm oak floors, lime plaster and low tungsten lighting'. You are trading praise the model can't act on for specifics it can.

Think of it as the difference between an adjective of judgement and an adjective of fact. 'Elegant', 'striking', 'premium', 'tasteful' are judgements - they describe your reaction to a thing, not the thing, and the model has no shared definition to act on because the training data attached them to everything from watches to wallpaper. 'Fluted', 'cantilevered', 'lime-washed', 'north-facing' are facts - they describe the thing itself, and the model learned a consistent look for each. The whole move of this lesson is learning to hear the difference in your own drafts and swap judgement for fact every time.

The corollary is liberating: you don't need flowery language at all. A design prompt reads more like a materials schedule than a poem, and that is exactly why architects and designers - who already think in named systems - have an edge here.

It helps to keep a mental map of the families of terms, because each fills a different slot from Lesson 1.1.

Movements and periods are among the strongest steers: 'brutalist', 'art deco', 'bauhaus', 'metabolist', 'deconstructivist', 'Palladian', 'high-tech'. They carry whole grammars of form. Named styles and idioms are gentler but precise: 'Japandi', 'wabi-sabi', 'mid-century modern', 'Kerala vernacular', 'Mediterranean', 'industrial loft'. Materials are reliably powerful because they are concrete: 'travertine', 'laterite', 'board-formed concrete', 'Corten steel', 'terrazzo', 'burnished brass', 'lime plaster', 'teak'.

Light terms borrow from photography and set mood: 'golden hour', 'overcast', 'raking light', 'chiaroscuro', 'backlit', 'soft north light'. Medium and render words decide the picture's whole nature: 'architectural photograph', 'axonometric', 'watercolour', 'gouache', 'blueprint', 'V-Ray', 'unreal engine render', 'physically based rendering'. Camera and lens words finish the framing: '24mm', 'tilt-shift', 'aerial', 'eye-level', 'shallow depth of field'. Learn a handful of live, high-signal terms in each family and you can compose almost any image.

You don't need hundreds of words per family - you need a reliable dozen you actually know. A working professional often runs on a surprisingly small core vocabulary, deployed with precision, plus the confidence to test a new term when a project demands it. Breadth is nice; fluency with a proven few is what ships work. That's why the personal-list habit below matters more than memorising any published 'prompt keyword' mega-list: a borrowed list is someone else's fluency, and it won't be there when the tool's autocomplete isn't.

There's a trap on the far side of specificity: reaching for a term so obscure the model never learned it. Naming a little-known regional architect or a boutique material may steer nothing, or worse, drift toward whatever few mislabelled images existed. The sweet spot is specific but well-represented - terms with enough presence in the training data to have formed a stable cluster.

This matters especially for Indian and other non-Western contexts, where the data is thinner and more stereotyped. 'Indian house' tends to collapse into cliche; 'Kerala vernacular with sloping Mangalore-tile roof and laterite walls' pulls far more faithfully because each component is well-attested. When a broad regional term underperforms, decompose it into its named material and formal parts - climate response, roof form, material, courtyard type - and let those specifics do the work the umbrella term couldn't. Module 6.4 goes deep on Indian-context interiors; the vocabulary habit you build here is what makes it possible.

When you're unsure whether a term carries weight, test it in isolation - one term, everything else held constant - and see if the image genuinely moves. That single-variable test (Lesson 1.4) is also how you discover vocabulary, and it's the only honest way to tell a term that truly steers from one you merely hope does. Guessing wastes rolls; testing builds a list you can trust.

The most valuable output of this module isn't a single image - it's a personal vocabulary: your own running list of terms that reliably did what you wanted, grouped by family. Professional AI users all keep one, because tool memory is short and yours isn't.

Start a simple document with columns for movement, style, material, light, medium and camera. Every time a term clearly earns its place - it steered the image the way you intended - write it down with a one-line note on what it does ('board-formed concrete -> visible timber grain in the concrete, reads as crafted, not raw'). Over weeks this becomes a palette you reach for without thinking, and it compounds: your prompts get faster and better because you're assembling from parts you've already proven.

Two refinements make the palette sharper. First, note antagonists - terms that fight each other ('minimalist' plus 'ornate' cancels out, 'photoreal' plus 'watercolour' produces mush) so you stop wasting rolls on contradictions. Second, note strength - which terms dominate a prompt even when buried, so you can place them deliberately. A curated vocabulary is the difference between prompting from memory and prompting from a proven instrument. It is also the most portable skill in this course: it survives every tool change, because it lives with you.

Individual words are only half the craft; how they combine is the other half. Terms stack, and they can reinforce or cancel. 'Brutalist' plus 'board-formed concrete' plus 'raking light' reinforce - they belong to the same world, so the image sharpens. 'Minimalist' plus 'ornate', or 'brutalist' plus 'delicate', pull in opposite directions, and the model resolves the fight by averaging toward a bland compromise that satisfies neither. Learning your vocabulary means learning which terms are friends and which are enemies, so you compose coherent worlds rather than contradictions.

There is also a hierarchy of dominance. Some terms are so strong they colour everything around them - name a famous movement or a signature material and it can override subtler cues you cared about. When a dominant term is drowning the rest of your intent, that's your cue to either move it later in the prompt, dial it down with a weight (Lesson 1.3), or split the idea across a couple of gentler terms.

And there is an honest ceiling worth stating plainly. Vocabulary steers appearance - it can make a facade read as travertine or a room as Japandi - but no word makes the model understand structure, program or code. A perfectly chosen prompt still produces appearance without logic (Module 0.1). Vocabulary is how you get the look exactly right; it is never how you get the building right. That distinction is what keeps a skilled prompter from mistaking a beautiful render for a resolved design - and it's the thread that runs through every remaining module of this course.