Prompts are config, not code

Manage LLM prompts like feature flags

Your LLM prompts are the highest-churn, most-experimented artifact in an AI app — so they belong in the same system as your flags and config. Store every prompt in Quonfig: version-controlled in git, changeable without a deploy, targetable by cohort, and type-safe. One targeting engine, one set of SDKs, one git-powered audit trail.

Store

Store prompts as config

A string or JSON config holds the prompt. Edit it in the UI, the CLI, or git — whichever fits your workflow. Every change is a real git commit with an author, a timestamp, and a diff, and the prompt changes without redeploying your app.

Change the prompt at runtime — no deploy, no release train
Every edit is a reviewable git commit (diff + author + timestamp)
Roll back to any prior prompt the way you roll back any commit

configs/support.triage-prompt.json

1{
2  "key": "support.triage-prompt",
3  "type": "config",
4  "valueType": "string",
5  "default": {
6    "rules": [{
7      "criteria": [{ "operator": "ALWAYS_TRUE" }],
8      "value": {
9        "type": "string",
10        "value": "You are a support triage assistant. Classify the ticket from {{customer}} (plan: {{plan}}) into one of: billing, bug, feature, other. Respond with only the label."
11      }
12    }]
13  }
14}

edit-without-deploy

# Edit the prompt in the UI, the CLI, or git — your choice.
$ qfg set-default support.triage-prompt \
    --value "You are a support triage assistant. Classify the ticket..." \
    --environment production
# Every change is a real git commit: author, timestamp, diff.
$ git log --oneline -1 configs/support.triage-prompt.json
a1f3c2e  support.triage-prompt: tighten classification rubric

Type safety

Mustache templating, fully typed

Prompts are rarely static strings — they interpolate user, plan, and request data. Quonfig string configs support Mustache {{variables}}, and qfg generate turns each one into a checked function parameter.

qfg generate

# Generate type-safe accessors for your prompts (TypeScript: node-ts, react-ts).
$ qfg generate --targets node-ts
Note: generated code imports 'mustache' for templated config values.
      Add it to your project: npm install mustache @types/mustache

What you get

A render function per templated prompt, with every {{variable}} typed as a parameter
Miss a variable and the build fails — not the model call
TypeScript targets: node-ts and react-ts

quonfig.generated.ts

1// AUTOGENERATED by quonfig-cli's 'generate' command
2import Mustache from "mustache";
3
4// Each {{variable}} in the prompt becomes a checked parameter.
5// The accessor is typed as:
6//   supportTriagePrompt: (...params: [{ customer: string; plan: string }]) => string
7
8const qfg = new QuonfigTypesafeNode(quonfig);
9
10const prompt = qfg.supportTriagePrompt()({
11  customer: "Acme Corp",
12  plan: "enterprise",
13});
14// -> "You are a support triage assistant. Classify the ticket
15//     from Acme Corp (plan: enterprise) into one of: ..."
16
17// Forget a variable, or pass a number, and the compiler stops you
18// before the prompt ever reaches your model.

Accurate by design: typed generation is TypeScript-only (Node and React targets), Mustache variables are typed as string, and mustache is a peer dependency the generator reminds you to install.

Rollout

A/B test prompts with the same targeting engine

A new prompt is just a config change. Roll it out to 10% of users, target it to a cohort or environment, and roll it back instantly — driven by the exact same targeting engine that powers your feature flags.

Target by cohort

Serve a new prompt to enterprise users, one region, or a single environment. Same rule engine, same context, as your flags.

Percentage rollout

Ramp a prompt from 10% to 100% as confidence grows. Every step is a config change in git — reviewable and reversible.

Instant rollback

If a prompt underperforms, revert to the previous version from the UI or CLI. No deploy, no support ticket.

Proof

We dogfood this

Quonfig's own “infer a JSON Schema from examples” feature is powered by an Anthropic Claude prompt — and that prompt lives in a Quonfig config, not in our source code. When we tune the rubric, we edit a config and commit; no app redeploy. The feature that makes your prompts config is built on a prompt-as-config of our own.

our-config/configs/llm.infer-json-schema.json

A prompt change is a commit

We sharpen the schema-inference prompt by editing one config. The change ships through git history like everything else — author, timestamp, diff.

No redeploy to iterate

The app reads the prompt from config at runtime, so prompt iterations don't wait on a release. The same loop is available to you on day one.

Put your prompts where your config already lives

Version them in git, change them without a deploy, target them by cohort, and generate type-safe render functions. Get started in minutes.

Get Started