Deep Research to Skill

Turn research into candidate AI Agent Skills. Frame the question, compare sources, examine assumptions, and carry the evidence and open questions into review.

Follow the evidence. Preserve the reasoning.

The researcher

  • Matthew Clark: Research Scientist

Illustrative project: Recurring service alerts

Investigate what an AI agent needs to know before suggesting a response to a recurring equipment alert. Separate the observed signal from the conditions that could explain it.

Follow the workflow

  1. 1. Start with a decision worth investigating

    Matthew defines the question, the future agent task, and the limits of the investigation. A research brief identifies which conditions could change the answer and what evidence would make a candidate rule worth reviewing.

    A question worth asking: What would an agent need to know before distinguishing a recurring fault from a change in operating context?

    Research brief

    Investigate recurring alerts across operating states and equipment revisions. Identify the context an agent should request, the sources to examine, and the decisions that remain with a responsible expert.

    Before moving on: Can the question be investigated through specific cases and sources? Keep the intended use and the limits explicit.

    Explore AI Skill Engineering
  2. 2. Give each source a place in the investigation

    Bring together procedures, field notes, prior sessions, and reference material. Keep each passage connected to its origin and record the version, operating conditions, and limitations needed to interpret it.

    A question worth asking: Do these sources describe the same equipment, operating state, and version of the procedure?

    Source map

    A procedure describes how recurring alerts are reviewed during steady operation. A field note records a similar alert during a planned configuration change. Their contexts differ, so the observations need comparison before they support a common rule.

    Before moving on: Can a reviewer locate the original passage and understand the conditions it describes? Mark missing context as an open question.

    Explore Knowledge Source Intake
  3. 3. Test the explanation against the evidence

    Use AI-assisted exploration to compare passages, surface competing explanations, and form follow-up questions. Matthew checks the sources and records why an interpretation holds, where it depends on an assumption, and what could overturn it.

    A question worth asking: What evidence would distinguish a persistent issue from an alert associated with a planned change?

    Reasoning record

    The sources suggest that operating context matters, but they do not establish that a planned change explains every recurring alert. Preserve that distinction and seek a comparable case before proposing a response rule.

    Before moving on: Separate observations from interpretations. Record contradictions and missing evidence instead of smoothing them into one confident answer.

    Explore Reasoning Analysis
  4. 4. Make the findings ready for review

    Compile the investigation into structured skill inputs. Connect candidate decision criteria to their sources, preserve alternative explanations, and identify the unresolved questions that need expert review before the skill moves toward approval.

    A question worth asking: Which findings are supported well enough to examine, and which decisions still need another source or an expert explanation?

    Candidate skill package

    Recurring service alerts ยท research draft. Includes the research brief, source map, candidate context checks, reasoning record, contrasting cases, and open questions. Next step: AI Skill Validation. Publication requires its own review and owner approval.

    Before moving on: A reviewer can trace each candidate criterion to evidence and see where it does not yet apply. Keep the research draft distinct from an approved skill.

    Explore AI Skill Validation

Keep the evidence, inference, and unknown apart

In this illustrative comparison, a procedure describes recurring alerts during steady operation. A field note records a similar alert during a planned configuration change but does not establish whether the alert persisted afterward.

Evidence: The sources describe different operating conditions. Preserve both passages and their context.

Inference: Operating context may change how an alert should be interpreted. This remains a candidate explanation to examine.

Unknown: Were the equipment revisions the same, and did the alert continue after the change? Seek a comparable case or an expert explanation before proposing a rule.

Keep the investigation useful at the next handoff

Carry the evidence, candidate criteria, and limits into validation. Resolve open questions with an expert and exercise the proposed reasoning in representative cases before approval or publication.

Explore AI Skill Validation

RapidKT Plain Text Website