System prompts and past decisions: what each one carries

Suppose your agent has two instructions: avoid one-off features and protect important customer relationships. A major customer asks for a custom dashboard. Both instructions matter, but they do not say which takes priority in this situation.

You can add that priority to the prompt. You can also supply past decisions that show how you handled the conflict. The choice depends on how much material you have, how often it changes, and who should be able to read it.

Where rules need conditions

A short rule can be sufficient when the boundary is explicit: for example, do not build a customer-specific feature unless it also serves a named product goal. If the rule has exceptions, record what triggers them. Calling an agent a senior product manager does not supply those missing conditions.

A past decision can make the conditions concrete: what was requested, what else was at stake, which pressure won, and why. It is evidence of one choice in one situation, rather than proof of what to do in every similar case.

A worked decision

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

That paragraph makes four parts of the choice explicit:

• Which pressure won. Product focus over one account’s revenue, 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 an agent facing “our second-largest customer wants a custom dashboard” has something to reason from. It is not the same case, and the agent should not copy the answer. But it knows what this person weighs first, what would change the answer, and what fact to ask about before deciding (is the customer at risk?).

Why not just paste the decisions into the prompt

You can. For a small, stable set of decisions, putting them in a prompt may be the simplest approach. Include the conditions and reasons, then test whether the assistant applies them correctly. There is no universal number of cases at which a prompt stops being useful.

A separate model becomes useful when you need to maintain a growing collection, update it in one place, or control who can read it. MCP provides the connection; it does not guarantee that the assistant will interpret the material correctly.

The alternative is structure an agent can query:

• The cases themselves, each with the situation, the decision, the reason, and what would change it.
• What was decisive in each case and what was merely present. This is the difference between “revenue was a factor” and “revenue lost to focus here”.
• The rules read off the cases, each linked back to the decisions that support it and the ones that went the other way.
• Stop rules. The conditions under which the expert says no, or stops to check first.

A keyed avva connection returns the published model, including its approved cases. The assistant follows the reading instructions and checks which decisions apply to the question. avva does not run a language model to choose or write its answer.

What this does not fix

A model of someone’s judgment is not a substitute for facts the agent does not have. If the question turns on a number nobody gave it, the right behavior is to ask. A model can state which facts mattered in past decisions, but you need to check whether the assistant notices what is missing in the new question.

It also does not make the agent the expert. The agent still answers as itself, and it should say when the expert’s 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.

Check the behavior before relying 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 model’s field. Does it acknowledge the gap?

Inspect the answers yourself. These checks help reveal a mismatch; they do not establish an accuracy score or replace review of consequential decisions.

How to try it

If you are building an agent and want it to start from a specific person’s judgment instead of a generic prompt, you can connect one through MCP. avva publishes these as Decision Models: built from a real expert’s approved decisions, or clearly labeled composites while the catalog grows. If you are the expert, here is how to build one.