Our summary
Client requests in Notion need matching to GitHub issues each week in Akshay Pachaar's follow-up agent. TrueForge organizes the agent's work. A script shortlists candidates, then Jev selects a match or “none”, avoiding a text-writing model's judgment on every candidate.
There are ten issues in the shortlist. A confidence score estimates how sure Jev's selection looks, and Claude reviews scores below 0.80, or 0.90 for closed issues. Claude also drafts new issues when needed. These cutoffs route review; a person must still approve changes to Notion or GitHub.
Both versions matched thirty-seven of forty requests correctly in Pachaar's test. Adding Jev routed sixteen cases to review, where five errors were caught. His estimated costs favor Jev. That small author-run test does not establish the outcome for another backlog or another set of client requests.
Key takeaways
- Make a shortlist, then let Jev select one match or none.
- Give uncertain matches a second review; this example applies a stricter cutoff to closed issues.
- A confidence score does not replace human approval before writing to business tools.