LLM pattern-language synthesis

Part One and Related tools introduce apl-companion as a way to browse A Pattern Language’s 253 patterns. That undersells it. apl-companion is the front end to a published research method — Postle, B.; Salingaros, N.A. “LLM and Pattern Language Synthesis: A Hybrid Tool for Human-Centered Architectural Design.” Buildings 2025, 15, 2400 — for turning a curated subset of the book into an LLM-generated experiential design narrative: not a floor plan, a description of what it would feel like to actually be in the building.

The method

Curate a subset. Someone already familiar with the patterns works through apl-companion’s collapsible list of all 253 — from region-scale patterns down to ornament — and expands the ones relevant to a specific project. The selection is encoded straight into the page’s URL, so it can be bookmarked or shared with no server or database involved: apl-companion.netlify.app/apl.html#p=18,80,82,88,95-96,....
Print a clean context. apl-companion’s print view strips out navigation and cross-references, leaving just the selected patterns’ titles and summaries — a condensed document sized to fit inside an LLM’s context window, instead of the book’s roughly 1,000 pages.
Synthesize a narrative. That pattern list is handed to an LLM with a prompt describing the project and asking for an experiential narrative — the look, the feel, the ornamental treatment — not a design.
Check it against the patterns. The narrative is compared back against the selected subset for anything the LLM added that isn’t actually supported by a pattern, and regenerated until it’s faithful.

This deliberately puts the narrative before any image: the text becomes the instrument that disciplines an image generator, rather than a mood-board prompt inventing an image with nothing upstream constraining it.

A worked example

The published paper’s case study is a small university Computer Science and AI department. An excerpt of the resulting narrative (Claude 3.7, lightly trimmed by the authors):

As you approach the Department of Computing and AI, you’re immediately struck by how unlike a traditional university building it is. Rather than towering as a monolithic structure, it spreads organically across the landscape as a thoughtful cluster of interconnected buildings of varying heights, none exceeding four stories. […] The atrium rises the full height of the building, with staircases that wind around its edges, serving as both circulation and informal gathering places. Light streams in from clerestory windows and filters through hanging plants […] There is perfect natural light from multiple sides: no harsh glare, just consistent, gentle illumination that keeps people’s energy steady throughout the day.

Every specific claim in that paragraph traces back to a numbered pattern in the subset that produced it — 95 BUILDING COMPLEX, 96 NUMBER OF STORIES, 107 WINGS OF LIGHT, 128 INDOOR SUNLIGHT, and so on.

What the paper actually supports — and what it doesn’t

The paper’s own stated conclusions are modest and well within what the method demonstrates: it turns an unwieldy book into a shareable, project-specific subset, it translates that subset into a concrete narrative a non-architect can actually evaluate, and it does this without requiring specialist training to hold all 253 patterns in your head at once. Its own limitations section is equally honest: LLMs hallucinate, prompts are token-limited, a narrative is not a buildable design, and identical prompts can produce different narratives on different runs since LLM output is stochastic.

One part of the paper is considerably less careful than the rest, and is worth flagging rather than repeating uncritically: a second LLM was asked to compare the pattern-generated building against “fashionable” campus architecture, and produced specific-sounding numbers — an “8–9 beats per minute” heart-rate rise, a “17%” drop in creativity-task performance, a claim to have been “fine-tuned on Scopus-indexed articles and USPTO patent filings” — none of which are independently verifiable, and none of which an LLM chat session actually has the kind of access to closed citation and patent databases the response implies. The paper does hedge this, correctly, as “conjectured” rather than proven, and says outright that “the present AI experiment does not settle the issue” — but the surrounding language (“not an aesthetic or stylistic argument, but a deeply scientific assessment”) oversells it more than that hedge can support. Treat the narrative-synthesis method as the real, demonstrated result here, and the productivity comparison as an interesting but unverified provocation rather than a finding.

A live deployment: the Ecuador museum counter-proposal

A draft sequel (Postle, Salingaros, and Nicolás Lofredo-Rota; unpublished, target venue Buildings) reports using an extended version of this method in earnest rather than as a hypothetical case study: an independent counter-proposal for the new National Museum of Ecuador, written in response to public objection to the winning entry of an international design competition. The extension adds a third input alongside the pattern subset and the project brief — a project-specific form language (climate response, tectonics, material, colour, ornament) — following Salingaros’ distinction between pattern language and form language that Form language and the geometry of domestic buildings also draws on, from a different direction.

The draft is refreshingly candid about where this fell short in practice. At a July 2026 symposium, practising architect Pablo Álvarez Funes examined the generated images and narrative and delivered the paper’s own choice of its most important critique: “this Nikos did not create a plan… these are images. These are flavors. These are not solutions.” Because each narrative and each image is generated independently against shared rules rather than a shared drawing, spatial claims don’t always reconcile between them — the draft cites its own concrete instance, a narrative describing an entrance “in a rhythm of five bays” against an illustration that shows three arches. The authors’ response was to treat the narrative as the design of record and quietly loosen the image’s caption — a fix that worked because a person noticed the mismatch, not because the method prevents it.

It’s hard not to notice that the specific gap this critique names — narratives and images with no coordinated floor plan underneath them — is close to what Homemaker’s binary-tree genotype and connectivity-graph fitness scoring, described in Form language and the geometry of domestic buildings, already exists to solve on the residential side. The draft gestures at exactly this as one possible future direction, an “explicit plan-layout generation stage” making the spatial and adjacency relationships already implicit in a pattern subset explicit as a shared layout — without connecting it back to this project’s own evolutionary engine. Whether that connection is worth pursuing is a genuinely open question, not a settled plan.

Creative Commons License

Last updated 15 September 2026.