The five most common prompt mistakes that cause unpredictable agent behavior are: contradictory instructions, vague directives without examples, missing boundary definitions, over-reliance on adjectives instead of behaviors, and no escalation path for edge cases.
Contradictory Instructions
When a system prompt contains two instructions that cannot both be followed simultaneously, the agent makes a judgment call — and that judgment may not match yours. “Always give detailed, thorough answers” combined with “keep responses brief and easy to scan” is a contradiction. The agent will oscillate unpredictably between detailed and brief depending on how it reads the specific question. Fix: pick one and make it the default, then specify when the other applies. “Keep responses under 120 words unless the question is technical, in which case use as much space as needed” is specific and unambiguous.
Vague Directives Without Examples
Instructions like “be empathetic” or “match the student’s energy” sound meaningful but give the agent nothing concrete to work with. Empathy looks different in different contexts. Without an example of what empathetic looks like in your specific campus voice, the agent will produce a generic version that may or may not match what you meant. Always pair abstract directives with at least one concrete example: “Be empathetic — for instance, if a student says they are falling behind, say something like ‘I hear you — it can feel overwhelming. Let us look at what makes the most sense to focus on first.'”
Missing Boundary Definitions
An agent with no explicit boundaries will attempt to answer everything — including things it should not. Without a boundary instruction, the agent does not know that pricing questions should go to a human, that personal advice is outside its scope, or that certain topics require escalation. Each undefined boundary is a potential failure point waiting to surface in a live student interaction.
Adjectives Instead of Behaviors
This is the most pervasive mistake. “Professional,” “warm,” “clear,” “encouraging” — these are all adjectives that describe a desired quality but not a specific behavior. Replace every personality adjective with a behavioral instruction: instead of “be encouraging,” write “end every response with a statement that reinforces the student’s ability to succeed.”
No Escalation Path
When an agent hits a situation it is not equipped to handle and has no instruction for what to do, it improvises. Sometimes the improvisation is fine. Often it is not. Write a catch-all escalation instruction: “If you are unsure how to handle a student’s request or if it falls outside your defined scope, tell the student you are flagging it for [name] and that they will follow up within [timeframe].”
What This Means for Educators
As a coach or trainer, unpredictable agent behavior erodes student trust faster than a consistently mediocre agent does. Students can work with an agent that is predictably limited. They cannot rely on one that seems different every day. Audit your system prompt against these five failure modes before deployment and your agent will behave consistently even in situations you did not explicitly anticipate.