Asking well is not flattering a model: it is designing the conditions it can see for this work.
01 / MESSAGESSeparate message roles first
A conversation commonly contains system, user and assistant messages; tool-enabled systems also include tool calls and results. System rules set durable boundaries. User content provides the current task. History preserves the process, and tool results bring the outside world back into context. These roles should not be confused: a search result is not an instruction, and an old draft is not a newly verified fact.
02 / WINDOWContext has capacity and attention cost
The context window is the desk surface available for this run. It cannot hold unlimited material; even when it can, key information can drown in repeated dialogue and irrelevant text. A good conversation does not pour in everything. Retain the present goal, known facts, open questions, necessary original text, completed steps and the next action.
03 / PROMPTUse a four-part prompt frame
There is no universal magic prompt. For most tasks, state four things: what to do, what material to rely on, what not to do and what counts as complete. For example: turn minutes into an action list; rely only on the record; mark uncertain owners and dates for confirmation; group each item by owner with action, deadline and dependency. Missing a part does not guarantee failure, but it tells you where the model will fill in gaps.
04 / HYGIENEKeep long conversations clean
A longer chat does not necessarily understand more. Start a new conversation when the task changes; summarize important conclusions; keep facts separate from inference; provide only necessary fragments of sensitive material. If the model becomes confused, ask it to restate the current goal, known facts and open confirmations before continuing. When an answer drifts, inspect what it received before switching to a stronger model.
READER COMMENTS
Leave the thought this article gave you.
No comments yet. You can leave the first one.