System prompts and past decisions: what each one carries

Your agent has two rules: avoid one-off features, and protect important customer relationships. Then your largest customer asks for a custom dashboard. Both rules apply, and they point in opposite directions.

The prompt does not say which rule gives way. You know, because you have made this call before. This article is about getting that knowledge to the agent, and about when a prompt is enough for it.

What a system prompt does well

• Rules with a clear edge. “Never quote a price without the current rate card” needs no example.
• Format and tone. How long, how formal, in what order.
• Scope. What the agent handles, and what it hands to a person.

If your rules never collide, stop here. A prompt is the simplest tool, and you can read all of it in a minute.

Where a list of rules goes quiet

Rules collide in the cases that matter. You can add a priority to the prompt: “focus comes before revenue”. Now the agent refuses the customer who was about to leave, which you would not have done.

Each fix needs its own exception, and the exceptions live in your memory of specific cases. A job title in the first line does not supply them. Studies of persona prompts found no gain in accuracy from one on factual questions.

What one past decision carries

Take an example decision, illustrative and not from a real client, written the way an expert would explain it to a new colleague:

Four sentences, and four things no rule in the prompt said:

• Which pressure gave way. One account’s revenue, to product focus, in this situation.
• The threshold. Three customers, not one.
• The exception. Imminent churn changes the answer.
• The reasoning. It can be applied to a request the expert never saw.

Now take the dashboard request. It is a different case, and the agent should not copy the old answer. But it knows what this person weighs first, what would change the answer, and which fact to ask about: is the customer at risk?

Why not just paste the decisions into the prompt

You can, and for a small, stable set you should. Include the reasons and the conditions, then test what the assistant does with them. No fixed number of cases marks where a prompt stops working.

The decisions want a home of their own when one of these becomes true:

• The set keeps growing. Every pasted case is loaded on every turn, whether the question needs it or not.
• It changes. A prompt copied into five agents is five copies to update.
• Other people should read it. A prompt has no way to say who may read it, or whose decisions these are.

Outside the prompt, the decisions can be kept as something an agent queries. Each case keeps its situation, decision and reason, and what would change it. Rules read off the cases link back to the decisions that support them and the ones that went the other way. Skill, system prompt or MCP compares the places to keep them.

That is what a Decision Model on avva is. A keyed connection returns the published model, including its approved cases, and the assistant checks which decisions apply to the question. avva does not run a language model to choose or write the answer.

What this does not fix

Someone’s judgment is no substitute for facts the agent does not have. If the question turns on a number nobody gave it, the right behavior is to ask.

It also does not make the agent the expert. The agent answers as itself, and it should say when the record has no position on the question in front of it. That is more useful than a confident answer borrowed from a case that does not fit.

Four checks before you rely on it

• Ask about a known decision. Can the assistant explain its reason and conditions?
• Change one condition. Does it notice that the old decision may no longer apply?
• Leave out a decisive fact. Does it ask, or invent an answer?
• Ask outside the field. Does it say the record is silent?

Read the answers yourself. The checks show a mismatch when there is one. They are not an accuracy score, and consequential decisions still need a person.

How to try it

To see an agent start from one person’s decisions, open a Decision Model and put a question to it. No key is needed for a first try. Each is built from an expert’s approved decisions, or marked Demo while the catalog grows.

If the decisions are yours, start with a decision log, or build a model from your history.