Write down how the agent should work.

An agent that starts every conversation blank forces people to repeat operational knowledge and past decisions. A growing system prompt can hold some of that information, but it mixes stable identity, temporary events, tool configuration and rules learned from failure. When something goes wrong, it becomes difficult to see which instruction mattered or why it was added.

We wanted the same clarity that versioned configuration brings to infrastructure. A change should have a history. Stable knowledge should have a home. A correction should retain the evidence that taught it, rather than becoming another unexplained sentence in an ever-longer prompt.

We wrote it for teams moving from occasional AI assistance to agents that participate in ongoing operations and software delivery. They need a shared way to express what an agent should know, how it should communicate and where its authority ends. It also gives reviewers something concrete to inspect when behavior changes.

We published a methodology, file specifications, examples and a reference implementation. The structure separates identity and values, human context, tool manifests, curated long-term references, daily records, operational skills and scheduled checks. Operator Kit turns that structure into an installable scaffold. Our contribution is the method and its working structure, used alongside existing runtime, search and voice tools.

One central addition is the instinct: a learned rule with a trigger, action, rationale, confidence and supporting events. A new pattern starts cautiously; repeated evidence can strengthen it and contradiction can weaken it. Durable lessons can move into reference material, while weak rules can be retired rather than accumulating forever.

A session records what happened in a daily log. Maintenance promotes useful knowledge into references and turns confirmed corrections into rules with evidence. Future sessions load the relevant context. Version control records how those files change, making the operator's working knowledge reviewable and portable between compatible models.

We also described memory canaries: facts planted in one session that another should not know, used to probe separation between agents. They are a diagnostic technique, not proof of complete isolation. An agent can guess, fabricate or withhold an answer, so a canary cannot replace access controls or trustworthy execution boundaries.

The difficult part is separating character, instruction and enforcement. A file can describe values and expected behavior, but it cannot guarantee that a model follows them. We present IPAC as published, evolving research with a working scaffold, rather than a finished solution to trust or alignment. Production coding still needs external permissions, review and delivery gates.

Maintenance matters just as much as capture. A rule that once prevented a mistake can become wrong when the system changes. We keep the method focused on evidence, pruning and drift detection so versioned context stays useful. Our delivery system complements that work by enforcing operational boundaries in tools, where a confident model cannot simply talk past them.