You can get a careful answer from an agent by asking carefully every time. That wears thin by the third day. Most AI apps let you keep instructions that apply to every conversation. A short page there does more than any single well-phrased question. Here is what belongs on it, in plain English.
Source and age
Every market figure comes from a tool call, and the answer names the tool. If no tool returned the figure, say so and give no figure. This one line removes the most common fault, which is a number recalled from training and presented as current. The fault itself is described in the number the model made up.
With every figure, give the time of the reading and its age in plain words. Say whether the time was measured, fetched or unknown. If the reading is delayed, say delayed. NoVo’s payloads carry as_of, age_seconds and as_of_kind on every reading, so the agent always has this to hand.
The what-it-is-not rule
Say what the number is not, wherever the tool’s description warns of a misreading. A row named after an index may be a future. A House disclosure is a filing and may be weeks behind the trade. A perp’s weekend move is a market price on another venue. A reader poll is a poll. These warnings live in the tool descriptions, as covered in the warnings written into a market tool. The rule tells the agent to carry them through.
The gap rules
Three short lines cover gaps. A field named as gated is withheld, and the answer says withheld. A tool that returns an error gets reported as an error, with its retry time if one came back. A ticker outside the coverage gets a plain no. The dealer map covers SPY, QQQ and IWM, and no other symbol should be answered from it.
Limits and layout
Do not predict. Do not advise a trade. If asked what to do, say the tools describe the present and do not answer that. The agent will drift toward a view if you leave this out, because a view makes a tidy ending.
Say how long an answer should be and in what order. One line per reading. Tool and age at the end of each line. A fixed layout lets you see at once when a line is missing.
Keep it short, then test it
Eight rules fit on a page. Eighty do not get followed. A model handles a few clear lines better than a long document, and you can check a short page against an answer in a minute. If a rule is being skipped, move it to the top and make it plainer.
After writing the page, put a question to the agent that the data cannot answer and see what comes back. A set of such questions is in testing an agent with questions the data cannot answer. Run them again whenever you change the page.
The rules above were written against the MCP & API. They carry over to any market server that stamps its readings.