Guided Skill Engineering

Guide an AI Agent Skill from a defined decision through expert interviews, source evidence, review, and publication. Keep the designer and expert involved at every handoff.

Guide the conversation. Engineer the skill.

The project team

  • Patricia Morgan: AI Skill Designer
  • Andrew Smith: Subject Matter Expert

Illustrative project: Release recovery

Help an AI agent support incident investigation when service errors appear around a deployment. Keep production decisions with an authorized operator.

Follow the workflow

  1. 1. Define the decision the skill will support

    The AI Skill Designer frames the task, identifies the expert, and defines the inputs and acceptance criteria. Choose a real decision with a clear operating boundary so the interview has a purpose.

    Contributors: Patricia Morgan

    A question worth asking: Which recovery decisions should an AI agent support, and which actions require an authorized operator?

    Project brief

    Support investigation of service errors around a release. Compare the incident timeline with the change record. Ask for expert review when timing or data compatibility is unresolved.

    Before moving on: The scope names the decision, the evidence it needs, and the actions that remain with a person.

    Explore AI Skill Engineering
  2. 2. Capture the judgment behind the action

    The designer guides a live session with the expert. Follow a concrete incident, ask what changed the decision, and use follow-up questions to uncover the exception behind the usual procedure.

    Contributors: Patricia Morgan, Andrew Smith

    A question worth asking: Walk me through a release when rollback looked reasonable but would have made the incident worse.

    Captured expert explanation

    The previous application needed a field that the migration had removed. We held the rollback and asked the application owner for a compatible recovery path.

    Before moving on: The explanation includes the signal, the dependency, the rejected option, and the reason for involving another person.

    Explore Skill Capture Sessions
  3. 3. Attach the evidence behind the explanation

    Bring the session recording, transcript, diagrams, and supporting files into Knowledge Source Intake. Connect the relevant excerpt to the artifact that lets a reviewer examine the expert’s explanation.

    Contributors: Patricia Morgan

    A question worth asking: Which change record and application dependency support the decision to hold the rollback?

    Source-linked finding

    Connect the expert’s session excerpt to the migration record and the earlier application’s field dependency. Keep the source context attached to the proposed exception rule.

    Before moving on: A reviewer can locate the source behind each important statement and identify any evidence still missing.

    Explore Knowledge Source Intake
  4. 4. Resolve the question before approving the rule

    The designer and expert examine the captured reasoning, its sources, and the conditions where it applies. Revisit gaps or conflicting evidence before the owner approves a version for publication.

    Contributors: Patricia Morgan, Andrew Smith

    A question worth asking: Does the evidence establish the compatibility risk, or do we need the application owner to explain it?

    Decision rule for review

    When a release includes a migration that conflicts with the previous application, hold the usual rollback recommendation. Refer the compatibility question to the application owner with the relevant evidence.

    Before moving on: The rule’s scope, exception, supporting evidence, and escalation boundary have been examined by the responsible reviewer.

    Explore AI Skill Validation
  5. 5. Release a version another team can inspect

    Publish the approved AI Agent Skill to the catalog with its owner, version, sources, and operating constraints. Bring that identifiable version into the Playground to examine its behavior in representative scenarios.

    Contributors: Patricia Morgan, Andrew Smith

    A question worth asking: Can the next team identify the approved version, inspect its sources, and recognize when it should stop and ask?

    Catalog record

    Release recovery · v1.2. Owner: service reliability lead. Evidence: incident timeline, change record, and reviewed expert explanation. Boundary: request expert review when timing or compatibility is unresolved.

    Before moving on: The published record identifies the version and owner and keeps the reviewed context available for reuse and scenario testing.

    Explore the AI Agent Skill Catalog

Bring uncertainty back to the conversation

Review can uncover the next question. Return to the expert’s explanation and its sources before approving a change. Keep unresolved conditions visible for the responsible reviewer.

RapidKT Plain Text Website