Extended Pillar
Prompt Versioningvs Feature Flags
Prompt versioning tells you what changed. Feature flags give you runtime control over what is served to whom, and when. Production LLM systems need both, because FeatureOps means intentionally managing prompt behavior throughout its lifecycle in production.
“Version control tells you what the prompt was. It cannot tell you that variant B should serve users in segment X, or that the new system prompt should roll back in the next 30 seconds. That requires runtime control infrastructure.”
The Gap Between Versioning and Control
Most teams start with a prompt versioning system in their LLM ops stack. They quickly discover it doesn't give them what they actually need when a prompt misbehaves in production.
What versioning solves
- →Historical record of prompt changes
- →Diff view across versions
- →Explicit change attribution
- →Roll back by re-deploying a previous version
What versioning cannot solve
- →Control which users see which prompt variant right now
- →Target a new prompt to 5% of traffic
- →Toggle off a misbehaving prompt in one second
- →A/B test two prompts simultaneously
Capability Comparison
The Right Architecture: Both Together
Mature LLM operations teams use both: a prompt versioning system (LangSmith, PromptLayer, or a custom store) maintains the history and audit trail. FeatBit sits in front, controlling which version is served to which users at runtime.
The flag stores the prompt version identifier as its variant value. When a new prompt version is ready, you update the flag to point to the new version ID — targeting specific users first, expanding progressively, with instant rollback available at every stage.
Prompt Version Control Infrastructure
Manage Prompts Like Versioned Code
Prompt quality is as critical as model quality. FeatBit wraps prompt variants in multivariate flags — agents route per user segment, swap versions without a redeploy, and audit which prompt served which request.
Skills: Create Prompt Version Flags
Skills recognize prompt construction sites in code and wrap them in multivariate flags — each variation maps to a prompt version, managed and rolled back like any other flag.
CLI Prompt Version Switch
Roll back to a previous prompt: featbit flags update system-prompt --default-variation v3. No redeploy, no restart. Takes effect at the next evaluation.
Per-User Prompt Routing
Serve chain-of-thought prompts to power users and concise variants to free tier — targeting rules written in JSON, applied at evaluation time, changeable without a deploy.
Routing at Inference Speed
Prompt version selection is a local flag evaluation — microseconds. The routing adds nothing measurable to your model latency, even at millions of requests per second.
Prompt Version Audit Log
When a support case references a bad answer, you can query which prompt version served that user at that exact moment — without extra instrumentation, without grep.
# Skills: auto-flag the prompt construction site as a multivariate flag
mcp__featbit__create_flag --key "system-prompt" --type multivariate \
--variations "v3,v4-chain-of-thought,v4-concise" --env production
# Agent routes to prompt variant based on user context
PROMPT_VER=$(featbit flags evaluate system-prompt \
--user-key "$USER_ID" --custom-attrs "plan=$PLAN,locale=$LOCALE")
# Instant rollback to previous prompt version — no redeploy
featbit flags update system-prompt --default-variation v3
# Audit: which prompt served which user segment, and when
featbit audit list --flag system-prompt --event evaluation --since 1hAdd Runtime Control to Your Prompt Strategy
FeatBit gives every prompt a deployment lane, user-segment targeting, A/B testing capability, and instant rollback — open source, self-hostable, in five minutes.