The short version

Production agents become easier to maintain when expertise, working material, and software interaction are separate layers. Anthropic’s pattern combines Skills for reusable procedures, Files for task-specific documents, and computer use for actions in an application or browser.

Each layer changes at a different pace. A procedure may be revised occasionally, files vary for every case, and interface steps change when software changes. Treating them as one giant prompt makes all three harder to reuse and test.

Three layers

Expertise: A Skill contains the stable method, terminology, checks, and output expectations for a class of work. It is the playbook, not the case record.

Files: Documents, evidence, intermediate results, and finished artifacts belong to the current task. They can persist without being pasted repeatedly into instructions.

Interface: Computer or browser use carries out actions in the approved software. This layer should expose only the actions the workflow needs and should return evidence of what happened.

A reusable flow

A well-structured task can follow skill, files, action, saved result. Claude loads the relevant procedure, reads the current material, checks whether required information is present, performs an authorized interface action, and saves the confirmation or output back with the task files.

For example, a claims workflow can keep the filing procedure in a Skill, the intake and supporting documents in Files, and portal submission in the interface layer. The confirmation is saved as a result rather than disappearing into the interaction history.

Why separation helps

When the procedure changes, the team updates one reusable Skill. When a new case begins, only its files change. When the application interface changes, the interaction layer can be repaired without rewriting domain guidance.

This separation also clarifies failure handling. Missing or contradictory evidence is a file problem and should be escalated. A changed button or failed submission is an interface problem. An incorrect sequence of checks may require a Skill update. Distinguishing the layer makes diagnosis faster.

Avoid the giant-prompt pattern

A long prompt that combines every policy, all case documents, and click-by-click instructions consumes context and couples unrelated concerns. It is difficult to version and encourages copying stale details between tasks.

Keep stable knowledge reusable, reference task material through the file layer, and let interface tools handle observable actions. Store outputs and confirmations where later work can find them.

Practical checklist

  • Put stable procedure and quality standards in a Skill.
  • Keep case-specific evidence and outputs in Files.
  • Use narrow interface actions for approved software steps.
  • Validate required inputs before acting.
  • Save confirmations and finished artifacts.
  • Escalate missing evidence rather than inventing it.
  • Diagnose failures by layer before changing instructions.

Try it

Choose a recurring workflow and divide it into three boxes: reusable expertise, task-specific files, and interface actions. Then write its sequence from selected Skill to saved evidence.