BLOCK 4 · AI GATEWAY (PRODUCTION) 3:58 – 4:08 · 10 min · 3L / 7HO · Lab

TOPIC 4.7 / 10 · HO #6 · VERIFIED OSS

Prompts as gateway assets

Named, versioned, callable by header. Same shape as Git for code — but for prompts. Decouples prompt iteration from app deploy cycles.

4.7

Prompt Repository — versioned assets

10 min 3L / 7HO 3:58 – 4:08

Slide 1 / 3 · Versioned assets

Prompts should be assets, not constants

Named prompts (customer-support-greeting, repo-briefing-system) — each gets a UUID
Versioned — every edit creates a new version with a commit message
Called by header (x-bf-prompt-id: <prompt-UUID>) — gateway injects the latest committed version
Optional x-bf-prompt-version: <N> — pin a specific version (rollback / canary)
Invalid IDs are logged as warnings, request still runs

Slide 2 / 3 · Why this separation matters

Decouple prompt iteration from deploy cycles

Before:

Marketing rewords the greeting → engineer redeploys the app

After:

Marketing rewords the greeting → new prompt version in the repo
App still calls x-bf-prompt-id: <UUID>
No redeploy. No release window.

Slide 3 / 3 · Eight steps to prompt versioned

HO #6 — Prompt Repository (7 min)

Sidebar → Prompt Repository
Create Prompt → name workshop-greeting
Add a version → system message: "You are a friendly Vue.js workshop assistant. Greet the user."
Commit message: v1: initial draft → save
Add another version with a tone change → commit v2: more casual → save
Copy the prompt UUID from the dashboard (you'll need it for the header — name ≠ UUID)
(Optional) curl ... -H "x-bf-prompt-id: <UUID-you-just-copied>" — see the stored prompt injected
Confirm both versions visible in the repo view

Speaker notes

Verified OSS (2026-05-19): Prompt Repository is part of OSS Bifrost — confirmed against the installed instance. Create, version, and invoke prompts via header are all OSS features. The richer capabilities in the ent-v1.3.9 changelog (Jinja2 variables, sessions, playground, deployment, folders) sit on top of the OSS workflow this beat teaches. The HO stays in.

Slide 1: "Prompts as versioned gateway assets. Named — every prompt has a human-readable name AND a UUID. Versioned — every edit creates a new version with a commit message. Called by header — your app sends x-bf-prompt-id: <UUID> and the gateway injects the latest committed version of that prompt into the request. Want to pin to an older version? x-bf-prompt-version: 3. Same shape as Git for code — the prompt has tags and history, and you call the one you want by id."

Slide 2 — why decouple: "Imagine your marketing team rewords the customer-support greeting. Without a Prompt Repository: they email engineering, engineering edits the constant in the app, engineering deploys, marketing sees the change in production three days later. With the repo: marketing's request becomes a new prompt version. The app still calls the same prompt id. No redeploy. No release window. Prompts iterate on their own cadence."

Slide 3 — HO: "Seven minutes. Prompt Repository → Create Prompt → name it workshop-greeting. Add a version with a system message — pick anything friendly. Commit it as v1. Then add a v2 with a slight tone change. Both versions stay accessible. Copy the prompt's UUID from the dashboard. Optionally curl the gateway with the prompt header to see the stored prompt injected — but that's gravy."

Demo piece — Attendees

Click sequence:

Sidebar → Prompt Repository
Create Prompt → form: name workshop-greeting
Add Version: messages array with one system message: "You are a friendly Vue.js workshop assistant. Greet the user warmly." Commit message: v1: initial draft. Save.
Add Version (again): same prompt with tone change: "yo! welcome to the workshop. let's go." Commit: v2: more casual. Save.
Confirm both versions appear in the version history.

Optional curl (instructor projector if time — replace <UUID> with the prompt's actual UUID from the dashboard):

curl -X POST localhost:8080/v1/chat/completions \
  -H "Authorization: Bearer $WORKSHOP_VK" \
  -H "x-bf-prompt-id: <UUID>" \
  -H "Content-Type: application/json" \
  -d '{"model":"anthropic/claude-sonnet-4-6","messages":[{"role":"user","content":"hi"}]}'

Response uses the stored prompt as the system message — no prompt text in the request body. To pin a specific version: add -H "x-bf-prompt-version: 1".

Triage note: Tier 2 cut drops the "use via header" step (saves ~3 min). Concept lands in the lecture without the live demo.