Assume you have an authorized transcript or set of notes. It may be short or long; the important question is whether it contains a decision you can turn into a bounded piece of work.
A summary can be useful. It just does not tell you what the next step is, what information is missing, or which actions need an owner to approve.
Building the Harness
Start with one model and one job. Use a second model only when it gives you a specific advantage—such as a separate critique—not because a “multi-LLM” setup sounds sophisticated.
The useful output is a structured workflow card: the goal, allowed inputs, step sequence, uncertainty, stop conditions, and the person who owns the consequential decision.
"Review this authorized transcript for one narrow job: [Topic]. Extract only recommendations supported by the material. Create a reviewable workflow card with: goal, allowed inputs, steps, unknowns, stop conditions, and owner approval points.
Do not send, publish, submit, change settings, or schedule anything. When the transcript is unclear, label the uncertainty instead of guessing."
The Workflow Card: From Talk to a Reviewable Method
The point is not to make the browser act on an unreviewed transcript. The point is to give the owner a compact artifact they can inspect before a browser tool receives even a safe task.
Here is a synthetic example of a workflow card for a follow-up draft:
{
"workflow_name": "Follow-up draft review",
"allowed_inputs": ["approved opportunity notes", "owner-approved contact record"],
"steps": [
{"action": "extract_relevant_facts", "source": "approved notes"},
{"action": "draft_email", "tone": "plain and specific"},
{"action": "flag_unknowns", "rule": "do not infer missing facts"}
],
"stop_conditions": ["contact details are unverified", "claim lacks source", "owner has not approved a send"],
"owner_handoff": "Review the draft and decide whether to send."
}A reusable skill makes the method visible. It does not make old advice current, and it does not eliminate the need to check live facts before a real action.
But why go through the effort of turning the transcript into this JSON file? Why not just pass the raw mentor transcript to the AI every single time you need to do the task?
Once a workflow is clear, you can measure whether its compact context is actually cheaper and easier to review than replaying the full transcript. Part 3 shows how to do that without inventing a benchmark.
Sources, examples, and test notes
- OpenAI, A practical guide to building agents — supports defining tools, guardrails, and human intervention before an agent takes high-risk actions. Source type: official documentation. Checked: 2026-08-02.
- NIST AI RMF Core — supports documenting task scope, knowledge limits, and oversight. Source type: standards guidance. Checked: 2026-08-02.
Example status: Synthetic. The JSON illustrates a review-only workflow; it has no live account access and cannot send or schedule a message.
Last reviewed: 2026-08-02.