I nearly sent an investor memo with a citation that did not exist. The figure looked precise, sat inside a clean sentence, and traced back to a source I had never read. It was a blend of three separate papers, stitched together and attributed as if it came from one. I caught it minutes before the memo went out. It was not the first time. An earlier tool had once handed me a full bibliography that was entirely invented.
That was the wake-up call. The problem was not the model. The problem was that I was using it on its default settings, and the default is a helpful assistant, not a careful one.
The fix took about thirty minutes. It is a one-time configuration you write into the assistant's custom-instructions field, the preferences that load into every conversation before you type anything. Done once, it changes the behaviour of every chat after it. This is the first move in a four-part system: configure the model, then protect your projects, then make it learn, then enforce the rules. Configuration is the foundation the other three build on.
The five rules
These are the non-negotiables. Each one closes a specific failure mode I hit repeatedly.
- Search before claiming. Verify current facts with a live search rather than reaching into stale training data. Most fabrication starts as a confident guess about something the model half-remembers. Forcing a search first removes the guess.
- Surface uncertainty, do not fabricate. When the model does not know, it should say so plainly. A visible gap is useful. An invented detail that fills the gap is a landmine you step on later.
- Challenge, do not validate. The assistant should push back on weak reasoning, not agree with it. I want a third reader who tells me where the argument breaks, not a second pair of hands that nods.
- Clarify before assuming. On anything ambiguous, ask a grounded question with two or three framed options, not an open-ended "what do you want?" A framed choice reads as a colleague thinking. An open question reads as a tool stalling.
- Flag risks proactively. Surface the risk I did not ask about. An advisor volunteers the thing you missed. An order-taker waits to be asked and lets you walk into it.
None of these are clever. They are the habits of a good analyst, written down and made standing instructions so the model applies them without being reminded each time.
The writing discipline
The second half of the configuration governs how the output reads. This part I lifted straight from consulting training, where the standard is bottom-line-up-front: the answer first, then the reasoning, then the alternatives, then the next step. Not a slow build to a conclusion buried on the last line.
Alongside that, a short list of bans that strip out the tells of machine-written text:
- Banned openers: "Certainly," "Great question," "Absolutely," "It's worth noting." Filler that delays the point.
- Banned vocabulary: "delve," "navigate the complexities," "robust," "leverage" as a verb, "comprehensive." Words that signal padding rather than substance.
- No meta-labels: no "The read:" or "Here's the thing" preambles, and no manufactured lists of three where two points would do.
- Punctuation precision: no dashes standing in for a pause, and ranges written with "to" rather than a stroke.
The effect is output that reads like it came from someone who respects your time. Direct, unhedged where the facts allow, honest where they do not.
Why it holds
The reason to put this in the preferences field rather than retyping it each session is simple: a rule you have to remember is a rule you will eventually skip. Loaded into the configuration, the discipline runs whether you are sharp that morning or rushing between meetings. You set it once and stop re-teaching the same lesson.
Thirty minutes of setup buys you an assistant that verifies before it claims, tells you when it is unsure, and writes like a professional. That is the difference between a chatbot and a thinking partner, and it is entirely a question of how you configure the thing before you start.