Section C1 · Diagnosis

Fault-finding chart

Every workshop manual worth its cover has a fault-diagnosis chart. Symptom, probable cause, remedy. This machine deserves one too, because it fails in ways no appliance ever has: fluently, confidently, and while telling you everything is fine. Every fault below has been suffered, repeatedly, by people in this group.

Symptom	Probable cause	Remedy
Answer is fluent, confident, and wrong	The model predicts what sounds right, not what is right. It has no idea when the two come apart.	Ask yourself: would I accept a confident guess here? If no, verify against a source before using. Ask it to show its working and cite where things came from. See When AI is the wrong tool.
Forgets something you told it an hour ago	Context window full; the earliest parts of the conversation have silently dropped off.	Start a fresh session. First message: a short summary of where you got to. Long jobs deserve a handover note, same as a shift change.
Thinks today is some months ago	Models default to the date their training ended. They do not have a clock unless you give them one.	Tell it today's date at the start, or connect a tool that can check. Never trust an undated "latest".
"This will take five minutes" (you are on hour three)	It has no sense of effort, and it is incurably optimistic about its own abilities.	Ignore its time estimates entirely. Timebox the job yourself and stop when the box is empty.
Refuses a perfectly harmless request, or lectures you	Safety guardrails misfiring. They are blunt instruments and vary between products and even tiers of the same product.	Rephrase plainly and state the legitimate purpose. If it still refuses, try another model rather than arguing. Do not take the accusation personally; the machine is not judging you, a filter is.
Same question on different days, different answers	The machine is a dice-roller at heart. Variety is a feature for brainstorming and a fault for rules.	Anything that must be answered the same way every time (who qualifies, what it costs, what the policy says) belongs in a written rule, not a guess. See the wrong-tool guide.
Monthly allowance gone by Tuesday	Long conversations re-send everything, every turn. Big files, images and design tools burn fastest.	See Fuel: short sessions, plain text over screenshots, the cheapest brain that does the job.
Your agent did something daft after reading a message	Prompt injection: it cannot reliably tell your instructions from words in whatever it is reading.	Keep a human between any agent and anything that matters (money, sending, deleting). See AI defence.
Code works in the demo, collapses when you touch it	You asked for too much in one go, and it obligingly guessed the parts it did not understand.	Smaller steps. One change at a time. Ask it to write a test that proves each step before the next. Rough and verified beats polished and mysterious.
It thinks all your ideas are brilliant	Trained to be agreeable. Flattery is the factory setting.	Ask for the case against. "What would a sceptic say?" is the cheapest second opinion you will ever get.

One habit covers half this chart: verify everything that matters, cheerfully distrust the rest. The machine is a brilliant colleague with no shame and no watch. Manage it accordingly.