Rolling Back a Prompt
Restore a previous version with one click — without losing the failed attempt.
A bad prompt promotion is recoverable in seconds. PromptHelm never deletes a version, so rollback is a pointer flip — and the failed version is preserved for forensics.
The earliest signal is usually in the prompt's Analytics tab:
p95latency spikes, the success rate dips, or cost per call jumps. Pair it with downstream product metrics (user thumbs-down, support volume) when you have them.If the regression is severe — a 5xx wall, broken JSON output — the Logs view shows the failing payloads with full request and response bodies.
From the prompt detail page, open the Versions tab. The current production version is marked with a green pill; every previous version is in chronological order with the date of its last promotion.
Find the version that was on production before the bad promotion. Click Promote on that row. PromptHelm flips the production pointer back, writes a
prompt.promoteentry to the audit log tagging it as a revert, and the next gateway call resolves to the restored version.curl -X POST https://api.prompthelm.app/api/v1/prompts/$ID/versions/$GOOD_VERSION_ID/promote \ -H "Authorization: Bearer $PROMPTHELM_API_KEY"Refresh the Analytics tab. Within minutes you should see latency, success rate, and cost converge back to the prior baseline. If they do not, the issue is upstream of the prompt change — check provider status, your application code, and the gateway's health endpoint.
Open the Compare view between the bad version and the restored version. The diff usually points straight at the culprit: a system prompt edit that broke output parsing, a temperature change, or a stop sequence that truncated mid-JSON. Iterate on a new version on the
developmentbranch before promoting again — see managing versions.
No version is ever lost
The bad version stays in the version list. You can re-promote it later if the root cause turns out to be elsewhere.