Prompt Operations
This page shows how to run prompts through their life cycle with the features ANTS Platform has today. Each step links to the page with the details.
ANTS Platform has no separate PromptOps SDK, prompt template engine, A/B traffic splitter or automatic prompt optimizer. The workflow below uses Prompts, the Playground, Experiments, LLM-as-a-Judge and Automations.
The workflow
Draft
Write the prompt in the Playground. Compare models and wording in up to 10 windows side by side. Attach a structured output schema or tools if your application needs them. Save the window with Save as prompt.
Version and label
Every save in Prompts creates a new version. Your application fetches the version labelled production, so a new version is not live until you move the label. Use other labels, such as staging, for other environments. Add a commit message to say what changed.
Test before release
Build a dataset from real traces and known hard cases. Run a prompt experiment for the current and the new version, with LLM-as-a-Judge evaluators attached. Compare the runs on scores, latency and cost.
Release
Move the production label to the new version, in the app or with update_prompt(name=..., version=..., new_labels=["production"]). Applications pick it up within their SDK cache lifetime, 60 seconds by default. See Prompts API & Caching.
Roll back
Move production back to the previous version. No deploy is needed.
Watch
Link generations to the prompt version by passing prompt= to the generation in the SDK. The prompt's Metrics tab then shows, per version, generation count, median latency, tokens and cost, and trace and generation scores.
Guard the release
- Protected labels. In Project Settings > Protected Prompt Labels, protect
productionso only users whose role can manage protected labels can move it in the app. See Protected labels. - Notifications. Create an automation rule on the prompt version event to post to Slack or Teams, or call a webhook, when a version is created, its labels change, or it is deleted. A webhook can trigger your own CI or deployment checks.
- Audit trail. Prompt changes made in the app and through the API are recorded in the audit log.
Comparing versions in production
ANTS Platform does not split traffic between prompt versions. To compare two versions on live traffic, choose the version in your own code, for example by user or request, and fetch it with get_prompt(name, label=...) or get_prompt(name, version=...). Link each generation to its prompt, then compare the versions on the Metrics tab or with scores.
Next steps
Planned capabilities are listed on the roadmap.