§Field notes
Why niche AI experts beat general-purpose AI for the questions only specialists can answer.
General-purpose AI is built to be useful everywhere, which is exactly the problem when the question lives in a long tail. A Bronze Age pedigree holder, a variegated Monstera Albo cutting with a tissue-type parent, an ADU setback check on an irregular parcel — these are the questions no generalist model handles well, because “general” is itself an averaging. Three niches Lorequill already runs — vintage comic grading, rare cultivar care, hyper-local zoning — prove the point better than any abstract argument can. Here is what the narrow desks do that the broad ones do not.
Depth you cannot average your way into.
General AI is trained on the open web, which is wide and shallow precisely where it matters: across the back catalogs of long-tail niches. The best AI for comic grading is not a slightly larger model — it is a desk that has actually read thousands of CGC and CBCS label notes, knows the difference between a Blue River pedigree stamp and a Connecticut pedigree stamp, and holds the working table that maps a 7.0 raw with off-white-to-white pages to current comps on the relevant run. You cannot bolt that onto a base model and call it depth. A narrow desk starts with vetted sources — Census pull figures, Heritage lot archives for the era in question, the active GPAnalysis reports — and keeps them current on a continuous schedule. The generalist averages across snippets it finds; the specialist owns the corpus. The gap between the two is not a few percentage points; it is the difference between an answer you trust and an answer you have to verify yourself before you act on it.
Accuracy beats breadth.
A 7.5 raw estimate is only useful if the desk admits uncertainty and shows the comps behind it. Calibration is the whole game in expertise work — and calibration is what broad models do worst. A generalist will tell you a corner lot has room for a 1,200 square-foot ADU with the same tone it uses to answer anything else, regardless of whether the figure is correct for your parcel. A specialist desk refuses to commit without citation, hedges at the right moments (off-white versus white pages, a 1.5 cent press versus a true color touch, a corner-lot setback versus an interior-lot setback), and never lets a confident wrong figure pass into a reply. The cost of confident wrong answers in expertise work is asymmetric: one bad grade estimate on a five-figure key can be the difference between a fair trade and a lasting loss. Calibration is not an extra feature; it is the principal value the desk sells. A generalist trades confidence for breadth; a specialist trades breadth for honest answers.
Stop re-explaining the cover date.
Every chat with a generalist model costs setup time. You re-establish the cover date, the pedigree stamp, the cultivar registry entry, the parcel ID, the lot dimensions — every conversation, from scratch. The desk that has already absorbed the niche absorbs that context once, and it stays absorbed. If your question is AI plant cultivar help for a variegated Monstera Albo cutting, the desk already knows which registry entry you mean, what tissue type the parent came from, and which propagation medium your local humidity tolerates. You do not explain what an Albo is; you ask whether the cutting is reverting or showing sun stress on the white half. Pre-scoping turns the desk from a chatbot that needs context into one that already has it.
Citation discipline when the stakes are real.
A specialist desk refuses to answer a claim of substance without a citation. A generalist often will not, or cites a source that does not actually back the figure in front of you. If your AI zoning questionis whether an irregular lot’s ADU setback allows a 14-foot rear-yard addition, the desk has to tell you the municipal code section it cited, the parcel-zone overlay, and any lot-coverage exceptions — not just a confident number. Source citations are not a formality. They are the only thing that turns an AI answer into evidence you can act on in front of a buyer, a permitting office, or a tradesperson who needs to know whether the figure is right. The generalist hands you a string; the specialist hands you a defensible position. That is the difference between an answer you would commit to and one you would preflight by hand. The narrow desk earns the trust the broad model cannot, because the citation is the product, not a footnote.
§Try a desk
The desks Lorequill runs cover the niches general AI averages over.
Put a real question in front of one — a holder photo, a parent-plant picture, a parcel ID — and ask the desk the way you would ask a specialist you would pay by the hour. Curious what is already live? The catalog lists each niche, what it covers, and what the desk will and will not claim without more context.
in-app · no signup to read · no mailto for the front door
Copy: Lorequill editorial · Filed under long-tail expertise / specialist desks