We generated ~32,000 self-contained build prompts for Midnight

Midnight is a zero-knowledge L1: private state stays on the user's device, public state lands on chain, and the bridge between them is a circuit you write in a language called Compact. It's genuinely interesting technology. It also has one of the harshest first hours I've met in web3.

Not because the concepts are hard. Because the environment is. A hackathon dev sits down with a good idea and spends the next four hours on:

• a package set where @midnight-ntwrk/midnight-js-*, the proof server Docker tag, the ledger, and the wallet SDK all have to agree on a version, or nothing works;
• a local proof server that needs Docker, which on Windows needs WSL2, which needs virtualization enabled in BIOS;
• WASM + top-level await + a missing Buffer polyfill, which together turn any SSR framework into a wall of stack traces;
• a testnet wallet with no tDUST and no obvious way to get any.

None of that is the idea. All of it is tax.

So we built Creative Midnight — a site whose entire job is to collapse that first hour into a copy-paste. This post is about how the prompt generator works, what the numbers actually are, and the failure modes we hit in the reference builds, with the fix for each.

What the site is

Three things, in order of usefulness:

1. 1,996 hackathon ideas. Ten creative disciplines — dance, music, visual art, video, photography, writing, film & animation, games, theater, fashion — each with a market anchor and a "quantum hook" (the private-state mechanic that makes ZK actually load-bearing rather than decorative). 996 of those are base ideas; the other 1,000 are agentic-commerce overlays (A2A/AP2 agent negotiation, UCP ZK-checkout, x402 paywalls with a mimic USDC), distributed across the same themes so you can filter within a discipline.

2. A build prompt per idea, per network. Not a stub — a multi-thousand-line, fully self-contained prompt that includes the pinned package set, the Compact toolchain commands, the Vite config, the wallet boot code, the deploy script, and a list of red flags. You pick an idea, pick a network target, pick your OS, copy, paste into an AI coding tool, and the environment tax is pre-paid.

3. The scar tissue. Guides for the wallet, the proof server, the local (Undeployed) stack, Fly.io hosting, Android via the Kuira SDK, and a Known Issues page that exists purely so nobody re-debugs what we already debugged.

The interesting part: one idea, many prompts

Each idea is a row in a per-theme JSON file:

export type Idea = {
  id: string;
  theme: string;
  title: string;
  pitch: string;
  subDiscipline: string;
  quantumHook: string;      // the private-state mechanic
  quantumRationale: string; // why ZK is load-bearing here
  tam: string; sam: string; som: string;
  protocol?: "a2a-ap2" | "ucp" | "x402"; // agentic overlay, optional
};

The prompt is not stored. It is composed, by a single function:

export function buildVariant(
  idea: Idea,
  theme: Theme,
  network: NetworkVariant,
  os: OSTarget = "macos",
): string

buildVariant assembles shared blocks — PACKAGES, TOOLCHAIN_BY_OS, VITE_CONFIG, MIDNIGHTJS_BOOT, SIGNING_STRATEGY, WALLET_CLI_BLOCK, REDFLAGS, plus protocol blocks for the agentic overlays — around the idea's own text. That means a lesson learned once gets written once and appears in every prompt on the site.

The multiplier is the network variant, because the network genuinely changes the architecture:

Five visible variants × 1,996 ideas = 9,980 prompts in the UI. (A Mainnet variant exists in the codebase and is hidden — independent devs shouldn't be shipping vibe-coded contracts to a live chain.)

The OS toggle (macOS / Windows / Linux) is deliberately not a UI multiplier: it swaps one block inside a prompt rather than creating a separate prompt. For the offline bundles, though, we do expand it — which is where the headline number comes from:

{
  "ideaCount": 1996,
  "networks": ["preview","preprod","undeployed","undeployed-fly","mainnet","undeployed-mobile"],
  "oses": ["macos","windows","linux"],
  "variantCount": 31936
}

31,936 prompt variants, ~2 GB of text, generated by a build script and downloadable as per-target .txt bundles so you can feed them to your own model.

One file pins every version

Every prompt quotes the Midnight support matrix, which is the actual source of truth for compatible versions. Rather than sprinkling version strings through 2,000 lines of template, they live in one module:

export const MIDNIGHT_MATRIX = {
  snapshotDate: "2026-07-23",
  compact: { language: "0.23", toolchain: "0.31.1", runtime: "0.16.0" },
  midnightJs: "4.1.1",
  dappConnectorApi: "4.0.1",
  walletSdk: "1.2.0",
  indexer: "4.3.3",
  proofServer: "8.1.0",
  localStack: { node: "0.22.5", indexer: "4.0.2", proofServer: "8.0.3" },
  ledgerV8: "8.1.0",
};

When the matrix moves, one edit propagates to ~10,000 prompts, the setup guides, and the downloadable bundles. Note localStack.proofServer is 8.0.3 while the public one is 8.1.0 — those are not interchangeable, and discovering that by trial and error costs an evening.

Don't ship 2 GB through a Worker

First version of the download page imported the generated bundles as modules. The site is deployed on Cloudflare Workers and immediately returned Error 1102: Worker exceeded resource limits — the isolate has a hard memory ceiling and a multi-hundred-megabyte string blows straight through it.

Fix: the bundles never enter the bundle. Each is uploaded to a CDN and the repo holds only a pointer:

import previewMacos from "../../public/llms-prompts-preview-macos.txt.asset.json";
// { "url": "https://cdn.../llms-prompts-preview-macos.txt", "size": 127837211 }

The page renders a link and a human-readable size. The Worker stays small.

The failure modes, and the fixes

These come from four reference builds (choreokits, midnightfireside, flymidnight, and an Android one), and each fix is now baked into the prompts.

The one worth repeating: proofs are slow. A k=14 circuit takes 30–120 seconds. Every prompt therefore mandates a visible "Proving… this can take 30–120s" state that keeps the UI usable. Designing for sub-second finality is the single most common way a Midnight demo ends up feeling broken when it's actually working fine.

Using it

1. Pick a discipline, then an idea (filter by protocol if you want an agentic overlay).
2. Pick a network tab. If you don't know, pick Undeployed (Local) — no faucet, no waiting, and the genesis wallet is already funded.
3. Pick your OS so the Docker/toolchain block matches your machine.
4. Copy the prompt into your AI coding tool. Run the two toolchain commands it gives you in a terminal (installing Compact and starting the proof server are the only things a browser can't do for you).
5. Ship.

There's also a downloadable skill file if you'd rather give your own agent the whole knowledge base instead of one prompt at a time.

What I'd tell anyone building on a young ZK chain

Write the environment down. Every hour we lost was recoverable exactly once — the second time it cost nothing, because the fix was in a template instead of in someone's memory. The prompt generator is, honestly, just a very elaborate way of refusing to debug the same thing twice.

Links: Creative Midnight · Midnight docs · support matrix

Built during the Creative AI & Quantum Hackathon organised by StreetKode Fam during Indian Krump Festival 14.