01 / FOR HUMANS
The thinking behind the workflow
The decisions, the wrong turn, and why we kept a human in the loop.
Start with the decision, not the tool
Our first brief was simple: “Research the market.” The agent returned a confident overview. It was readable, detailed, and almost impossible to act on. We hadn’t told it which decision the research needed to support.
So we worked backwards. We needed to choose three customer interviews for a new service. That changed the task from collecting everything to finding a few useful signals, with sources we could inspect.
“The agent gathers evidence.
A person makes the call.”
Make the boundaries visible
We gave the agent a narrow scope: one customer segment, public sources only, and a short evidence table. Every claim needed a link. Missing evidence was a finding, not an invitation to fill in the blanks.
The important constraint was the handoff. The agent could suggest interview candidates. A person would review the evidence and decide who to contact. Nothing would be sent automatically.
Test the handoff before scaling
We tried one candidate first. One source was out of date; another did not support the claim. A polished summary would have hidden both problems. The evidence table made them easy to spot.
Only after that review did we run the next two candidates. The lesson wasn’t to write a longer prompt. It was to make success observable, keep the first run small, and give uncertainty a place to go.
PAUSE & PUT IT TO WORK / 5 MIN
Rewrite one brief you used this week.
Before opening your agent, write three lines: the decision you need to make, the evidence you need, and the point where a person takes over.
Then compare with our example
Decision: Who should we interview first?
Evidence: One source-backed candidate in our target segment.
Boundary: Stop before outreach. Ask me to review.
02 / FOR AGENTS
Less guesswork. A better handoff.
Produce a source-backed shortlist for human review. Never contact candidates.
EXECUTION CHECKLIST
- Ask which decision the research supports and confirm the customer segment.
- Find one candidate using public sources. Attach a source URL to each claim.
- Flag missing or conflicting evidence instead of guessing.
- Stop for human review before expanding the shortlist. Do not send messages.
research-brief.mdA fill-in brief with an explicit review boundary.
# Research brief Decision to support: [your decision] Customer segment: [one segment] Sources: public, with URLs per claim First run: one candidate Uncertainty: flag missing evidence Stop: human review before expansion or outreach
review-checklist.mdFour checks before trusting the output.
# Review checklist - Does each claim have a working source? - Does the source actually support the claim? - Is missing or stale evidence flagged? - Has a human approved the next step?
Inline sample files only. No agent token, download tracking, or API resource is created.
Small brief. One test. Human review.
Try it on a real decision before adding more automation.
Back to the studio ↗