AI Skill
Engineering

Turn expert judgment into a clear plan for AI Agent Skills. Define the decisions, evidence, exceptions, and acceptance criteria before capture begins.

Engineer the reasoning behind the work.

  1. 01 Start with a decision
  2. 02 Make the judgment explicit
  3. 03 Plan for review

01 / Define the work

Start with a decision worth getting right

A useful skill starts with a specific piece of work. What should the agent recognize? What evidence should it examine? Which decision should it help someone make?

Build the project around that outcome. Define the situation, the people who understand it, and the limits of the agent’s role. This gives every topic, question, and capture session a purpose.

A running example

Investigating errors after a software release

Task
Help an incident owner decide what to investigate before considering a rollback.
Evidence
The release timeline, service errors, change records, and comparable incidents.
Boundary
Explain the recommendation. Leave production changes with the authorized operator.

02 / Map the reasoning

Find the conditions that change the answer

Experts recognize the detail that makes one case different from the last. Organize topics around those distinctions, then ask questions that expose the decision behind the action.

  1. 01

    What did you notice?

    Capture the signal, the comparison, and the context the expert checked first.

  2. 02

    What would change your mind?

    Ask for the exception, the conflicting evidence, and a case where the usual response failed.

  3. 03

    When should the agent stop?

    Define missing inputs and decisions that need someone with the authority to act.

Explore a decision

Same symptom. Different next step.

Investigate the release Investigate other changes Ask for missing evidence
SymptomCompare evidenceNext step

Errors begin after deployment

Check whether the changed component could explain the failures. Timing supports an investigation, but does not establish the cause.

Ask the expert

“What would you check before recommending a rollback?”

Errors begin before deployment

Look for earlier changes and other causes. Capture why the release timeline changes the investigation instead of treating every error spike the same way.

Ask the expert

“Which other changes would you investigate, and why?”

The timeline is incomplete

Request the missing timestamps or an incident owner’s review. Make uncertainty an explicit branch so the agent knows what evidence it needs next.

Ask the expert

“What is the minimum evidence needed to continue?”

An illustrative reasoning map. Each project defines its own rules and boundaries.

03 / Plan the capture

Give every conversation a job to do

Turn the reasoning map into a focused capture plan. Connect each topic to the right expert, the evidence to bring, and the questions that still need an answer.

A connected capture plan

Bring the context into the conversation

Expert experienceSource materialReal cases
RationaleExceptionsOpen questions
ContextFocused questionsSkill evidence

Keep the unanswered question connected

If an explanation lacks a source or leaves an exception unresolved, carry that gap into the next session. The project plan should make the next useful conversation obvious.

The topic

A decision to understand

Group related questions around a concrete judgment, such as distinguishing a release defect from an unrelated incident.

The contributor

The person with the context

Match topics to the experts who have made the decision. Ask them to bring the cases that reveal how they think.

The session

A realistic agenda

Organize questions, supporting sources, and expected capture time. Use AI-assisted questions and prompts written by the designer to pursue the missing detail.

Explore Skill Capture Sessions

04 / Set the criteria

Define what good reasoning must demonstrate

Set acceptance criteria while planning the skill. Reviewers need to know what evidence to expect, which scenarios to examine, and what should send the work back for another expert conversation.

Gaps become the next capture questions
Define criteriaCapture evidenceReview reasoningPrepare for reuse
01 / Evidence

Can the rule be explained?

Require the source, the expert’s rationale, and the conditions where a rule applies. Identify contradictions that need review.

AI Skill Validation
02 / Scenarios

Does it handle the exception?

Specify a routine case, a difficult variation, and a case with missing evidence. Define the expected reasoning and when to escalate.

AI Skill Playground
03 / Ownership

Who can approve its use?

Identify who reviews the skill, what must be resolved before publication, and who is responsible when the underlying work changes.

AI Agent Skill Catalog

Start with one decision

What does your best expert see that others miss?

Bring a workflow and a case that required judgment. We can help turn it into a focused skill project, from the first question to the criteria for review.

Next: Skill Capture Sessions
RapidKT Plain Text Website