Point the optimizer at real calls, let it diagnose what went wrong, and generate a targeted rewrite. Review the exact diff, then sync it to your test agent or live in Vapi.
The rewrite adds an explicit empathy step before explaining disputed charges, requires confirming resolution, and reads the SSN disclosure verbatim — the three gaps behind the failing calls.
Pick the calls that failed (or the ones that went well) and the optimizer finds the exact gaps to fix.
The rewrite addresses each diagnosed gap and is tuned to your model — not a blind regeneration.
See a word-level diff and the score lift before anything ships — then push to your test agent or Vapi.
Select up to five calls and the optimizer reads their transcripts and metrics to surface concrete gaps — each tied to a failing metric and the calls that prove it. Edit, drop, or add gaps to steer the rewrite.
Across 5 calls the agent stayed defensive, skipped a required disclosure, and never confirmed resolution.
Acknowledge the caller’s frustration before reasoning.
Read the SSN disclosure verbatim before account access.
Ask if the issue is resolved before closing the call.
Two ways to optimize
Start from your worst calls to patch a regression, or from your best calls to make the agent reliably reproduce them.
Learns from failures. Diagnoses what the instruction is missing on calls where metrics failed or were weak, then rewrites it to close those gaps — driven primarily by the failed metrics.
Learns from exemplars (good calls). Extracts the strong behaviours from both transcripts and metrics, and rewrites the instruction so the agent reliably reproduces them.
Every change is versioned, reversible, and one click from your live agent.
Each rewrite is committed as a labeled version with its diagnosis, diff, and status. See exactly which one is current on the agent — and reset to any earlier version in one click.
Approve a rewrite and it becomes current on your RubricHQ test agent. When you're ready, push the same prompt straight to your live agent in Vapi or Retell — or reject it and nothing changes.
How it works
Pick up to 5 calls. The optimizer reads them and surfaces the gaps to fix, each tied to a metric.
A targeted rewrite is generated to close each diagnosed gap, tuned to your model and constraints.
Read the word-level diff and score lift, then sync to your test agent — or push live to Vapi or Retell.
Diagnose from real calls, generate a targeted fix, and ship it with confidence.