Precedent & Case-Study Research

Recall the spine of this course: Claude is a plausibility engine, not a truth engine. Nowhere does that bite harder than in precedent research, because the output looks exactly like real scholarship. When you ask about a well-documented building - a Corbusier, a Correa, a Geoffrey Bawa house - Claude often does very well, because that information appeared thousands of times in its training data and the plausible answer is also the true one. But push toward the specific and the obscure - the exact built area of a lesser-known project, the client, the year, the structural span, a citation - and the ground gets soft. Claude does not know what it does not know; it will not pause and say "I am unsure of this figure." It will simply generate the most likely-looking number and move on, in the same confident voice it used for the facts it had right.

This is why the useful mental split is between retrieval and generation. Retrieval is looking a fact up in a source that actually contains it; generation is producing plausible text. Claude, on its own in a chat, is doing generation even when it feels like retrieval - it is reconstructing from patterns, not reading a record. The professional move is to supply the record yourself. When you paste a project monograph, an architect's own project description, a journal PDF, a competition report or your own site notes into the chat, Claude shifts from recalling to reading: it works from the text in front of it, and its claims become checkable against that text. This is the same idea the industry calls retrieval-augmented generation - ground the model on real documents rather than its memory - done by hand, in a chat window, by a designer.

So the first decision in any precedent session is not what to ask, but what to bring. If you have the sources, bring them and tell Claude to work only from what you provided. If you are genuinely exploring - "what typologies should I even be looking at?" - use Claude's memory deliberately, as a lead-generator, and mark everything it produces as unverified until you have found the primary source. Both modes are legitimate. The failure is confusing them: treating a memory-driven brainstorm as if it were a cited literature review, and letting invented specifics harden into "facts" that end up in a report.

In practice, grounding is a habit made of small, repeatable instructions. Start by feeding real material - a PDF of a project write-up, pasted text from an architect's website, a scanned award citation, your own visit notes - and then constrain Claude explicitly to it. A grounding prompt reads roughly like this:

Here are five project descriptions I have gathered (attached / pasted below).

Using ONLY these sources, build a comparison of how each handles
daylight and cross-ventilation. For every claim, quote the exact phrase
from the source it came from. If a source does not mention something,
write "not stated" - do not fill the gap from general knowledge.
List anything you are unsure about separately.

Three things in that prompt do the heavy lifting. "Using only these sources" tells Claude not to reach for its training data. "Quote the exact phrase" forces traceability - you can now check each claim against the text in seconds, and a claim with no quotable source is a red flag. "Not stated, do not fill the gap" is the single most valuable instruction in research work: it gives Claude explicit permission to leave a blank, which is what stops it inventing. LLMs hallucinate partly because the implicit reward in ordinary chat is to be helpful and complete; you have to actively license the honest "I don't have that."

When you genuinely need Claude to range beyond your sources - to suggest precedents you have not found yet - separate that request and label it. Ask for "candidate precedents to investigate, clearly marked as leads I must verify," and treat the output as a to-do list of things to look up, never as findings. A good discipline is to keep two visibly different buckets in your notes: Confirmed (traced to a primary source you have seen) and Leads (Claude suggested it; not yet verified). If your web search connector is enabled on your plan, Claude can help find the source - but the same rule holds: a link is a place to go read, not a licence to trust the summary. You still open it. Grounding is not a one-time setup; it is the posture you hold for the whole session, and the moment you relax it - "oh, just tell me roughly" - is the moment fabrication creeps back in.

Once your facts are grounded, Claude becomes genuinely excellent at the part of precedent work that is pure drudgery: making disparate case studies comparable. Ten projects described in ten different registers - one in an architect's poetic prose, one in a dry data sheet, one in your own shorthand - are hard to hold side by side in your head. Claude will normalise them into a single structure so the real comparison becomes visible. The instrument is a precedent matrix: a table where each row is a project and each column is a dimension you care about - typology, context, built area, structural system, key passive strategy, materials, what worked, what you would avoid, and crucially a Source column.

The skill is choosing the columns, because the columns encode your design question. If you are researching daylight in deep-plan offices, your columns are floor depth, section strategy, glazing ratio, shading, and measured or claimed daylight outcome - not a generic "description." Tell Claude the design question and ask it to propose the comparison dimensions first; refine them; then have it populate the matrix from your grounded sources. Add a final column headed "Verify" or "Confidence" and instruct Claude to flag any cell it could not source. That single column turns the table from a decorative summary into a working document: you know exactly which numbers still need a primary reference before they can go into a report or a client presentation.

Beyond the table, Claude is a strong synthesiser of patterns across precedents - "what do these six naturally-ventilated schools have in common, and where do they diverge?" - which is often the actual insight you were after. It can also translate a precedent into a lesson for your specific site: "given my brief and constraints, which two of these strategies transfer, and what breaks?" Here you are using Claude as a thinking foil, not an authority - the pattern it names is a hypothesis you test against your own judgement and the primary sources. Used this way, an afternoon yields a structured, sourced, comparable body of precedent that would have taken a junior a week - and every cell is traceable, which is the only kind of research worth putting your name to.

Assume, always, that some fabricated specifics are hiding in Claude's precedent output - and build the catch into your workflow rather than hoping to spot them by luck. The tells are learnable. Be most suspicious of precise quantities ("built area 4,270 sq m," "span of 18 m"), exact dates and named clients, and citations - book titles, page numbers, journal issues, awards. These are the coordinates Claude is most likely to smooth over with a plausible guess, because a specific-sounding number reads as authoritative. Fluency is not evidence: the most dangerous fabrications are the best-written ones. And a fabricated citation is uniquely damaging, because it looks like proof - a made-up book reference can sail into a dissertation or a competition essay and detonate later.

Verification does not have to be heavy; it has to be proportionate - the course's rule of scrutiny scaled to stakes. A precedent you mention in passing during an internal brainstorm needs a light touch. A precedent whose area or performance figure you are about to quote to a client, put in a report, or lean a design decision on gets checked against a primary source, full stop. A useful move is to make Claude audit its own output: paste its list back and ask, "For each project and each figure, tell me how confident you are and what the primary source would be - flag anything you may have inferred rather than known." It will often downgrade its own shakier claims when asked directly, which surfaces the soft spots for you to target. This is not because Claude "knows" it lied; it is because the question changes what a plausible response looks like.

The posture to carry out of this lesson is calm suspicion, not fear. Claude genuinely makes you a faster, broader, better-organised researcher - it reads more than you can, structures better than you will at 6pm, and never tires of one more comparison. What it cannot do is be accountable for the truth of a single claim. So you keep the two buckets - Confirmed and Leads - you demand quotable sources, you check the specifics that matter, and you treat every unsourced number as guilty until proven real. Do that, and precedent research becomes one of the highest-return uses of Claude in the whole of practice. Skip it, and you are quoting rumours in a nice font.