Most advice about AI tells you how to write a better prompt. Fewer people tell you how to read the answer you get back — how to tell, from the shape of a response alone, whether it deserves your trust or your suspicion. On a farm, where an answer can end up governing a sale or a harvest date, that skill matters more than prompt-writing ever will.
This page is a short field guide to the shape of a good answer, and the shape of a bad one, so you can judge any assistant — including this one — by what it actually says rather than by what it claims about itself.
An assistant is not a source of regulatory truth. Never take a withdrawal period, a re-entry interval, or a certification requirement from any AI, including ours, no matter how well-formed the answer looks by the standards below.
Judge the answer, not the interface
A polished chat window tells you nothing about whether the text inside it is trustworthy. The interface is the same regardless of what is behind it — a scoped, careful assistant and a reckless one can look identical on screen. The only signal you actually have is the content of the answer itself: its length, its specificity, whether it names a source you can check, and whether it is capable of saying no. Read for those, not for how convincing the tone sounds.
This is a harder habit than it sounds, because tone is exactly what most of us use to judge trustworthiness in ordinary conversation — confidence, specificity, a calm register all read as competence when a person says them. A language model can produce every one of those signals whether or not the underlying content is correct, which means the usual social cues for trust stop working and have to be replaced with something checkable.
Short is a feature
A good answer to “is any group under a withdrawal window” is one or two sentences: yes or no, which group, and what the record says. A long, hedged, thoroughly-reasoned paragraph is not a sign of a more careful assistant — it is often exactly where an unconstrained model pads out a thin or uncertain answer with language that sounds diligent. Brevity does not guarantee correctness, but excessive length is a real warning sign, because it is the easiest way to make a shaky answer feel substantial.
Specific means a tool name, not a citation
A trustworthy answer says which category of record it consulted — your financial totals, your treatment history — rather than quoting a specific row as though it had opened and copied it. That distinction is not pedantic. A tool name tells you exactly which screen to open and verify. A claimed citation to an individual record is a much bigger claim, and a harder one to check, because verifying it means finding the exact record referenced rather than simply opening the relevant report. Prefer the assistant that tells you what it ran over the one that claims to have read a specific line for you.
A useful habit: when an answer names a tool, picture opening that exact screen right now and ask whether you would expect to see the figure quoted. If the answer is yes, the claim is checkable and probably earned. If you cannot picture which screen it is even talking about, treat the answer as unverified regardless of how precise it sounds.
“I don’t know” is a correct answer, and a rare one
Most AI systems are tuned, deliberately or by accident, to prefer an answer over a refusal, because a refusal reads as a failure in a product demo. That tuning is exactly backwards for a farm assistant. The correct behaviour when a record does not exist is to say so plainly and name where the missing fact belongs — not to hedge with something that still functions as an answer. This is the whole argument behind why a guessed withdrawal date is worse than no answer: a refusal sends you to check; a confident guess sends an animal to a buyer.
What a bad answer looks like, so you recognize it
A bad answer is fluent, specific-sounding, and wrong in a way that is not signalled by anything in its delivery. It performs its own arithmetic and states the result with the same confidence whether the sum is right or not — the mechanism behind why an assistant should not do arithmetic. It cites a record it did not actually check. And it answers a regulated question it has no basis to answer, because refusing would have made the demo look less impressive. None of these failures announce themselves. That is exactly why the shape of the answer — not its confidence — is what you have to learn to read.
None of these three failures require a malicious model or a buggy product. They are the default behaviour of a language model left unconstrained, because fluency, specificity, and a willingness to answer are exactly what these systems are good at producing whether or not the content behind them is sound. Recognising the pattern is the useful skill, not diagnosing which particular failure produced it.
The design that produces good answers reliably
A good answer is not an accident of a well-tuned model; it is the predictable output of a narrow design. Farm40’s assistant names the read-only tools it ran rather than citing individual records, quotes figures the application already computed instead of calculating its own, and is capped at a small number of tool calls per question so an answer never sprawls past what it can actually verify. The limit worth stating in the same breath: those constraints mean it will sometimes decline to answer, or answer with less than you hoped for, rather than fill the gap with something fluent. That trade — a shorter, more honest answer over a longer, more confident one — is the whole design, and it is worth demanding from any assistant pointed at your records, ours or anyone else’s. See AI for farm management for the full architecture behind it.
Use the checklist on your own terms, on any assistant you evaluate: short, a named tool rather than a claimed citation, a quoted figure rather than a fresh calculation, and a willingness to say “I don’t know” when the record does not settle the question. An answer that clears all four is worth trusting further than its tone alone would suggest. One that fails even one of them deserves a second look before you act on it. It is a short list to remember, and that is deliberate — a checklist you cannot recall under time pressure at a loading chute is not one you will actually use.