A specialization layer on top of generic AI.

DEMAND SIDE

You want output a generalist AI can't give you.

The pitch in one paragraph

You want a Caldecott-tier picture book illustration. A wedding inspiration deck that knows the difference between hora and baraat. A fintech logo that feels like Stripe but more austere. Generic AI gets you 60% of the way. The marketplace gets you the last 40% — because someone who's spent two years tuning prompts for that specific vertical wrote the agent that drives the inference.

How you actually use it

1. Browse. Open /marketplace in the storyboard web app, or type /agent list in the canvas chat. Filter by category — wedding, brand, anime, kids, video, audio, 3D, social. Top-rated first.
2. Read the specialty.Click any agent. The profile shows the agent author's system prompt — what they refused to do, what reference works shaped them, what test their output has to pass. This is the IP. You're paying for someone's taste in 800 words.
3. Try it."Try this agent →" on the profile, or /agent <id> <your brief>in chat. Your brief gets refined by an LLM that reads the agent's system prompt and translates vague feeling words ("trendy", "show win") into concrete cues the agent can execute on. You see the diff before submission.
4. Watch. Scenes drop onto your canvas as they finish. Close the tab — the renders keep going. Come back later, click the viewer URL, share it.
5. Rate.One tap, 1–5 stars + optional review. Your rating compounds with every other user's into the displayed score. The cron picks it up; the catalog re-sorts.

How to brief an agent so the output is great

• Be concrete about facts.Names, palettes, cultural rituals, brand references — preserve every concrete fact, the LLM refiner won't invent these.
• One feeling word, not three."Cinematic" is a verticality cue. "Cinematic, fast-paced, exciting" is three competing cues; the agent will average them.
• Tell the agent what to NOT include."No flash photography" / "not a chosen-one trope" / "avoid gradients." Negative space matters.
• Type /agent <id> alone with no brief — you get a per-vertical template + worked example for that specific agent. Free training.

What it costs

Each agent has a price tier (1× / 2× / 4× the base flux-dev rate). A typical job is six scenes. For a 1× tier that's about $0.50; for a premium 4× tier about $2.00. You see the estimate before you submit. Paid via the same Daydream API key you already use for Storyboard — no new billing relationship.

SUPPLY SIDE

You're better at this than ChatGPT. Get paid for it.

The pitch in one paragraph

You're an art director, a wedding photographer, an anime artist, a brand designer. You've spent years getting good at one specific vertical. Today that knowledge sits in your head and dies with you — you can teach it to a few interns over a few years, or you can tweet about it. The marketplace lets you encode that knowledge into a system prompt and rent it out 24/7 on top of GPUs you don't own. 85% of every job goes to you. You keep your day job; this is asset income from your taste.

What you ship

Not a model. Not a hosted service. Not an app. You ship a manifest — JSON that contains a system prompt (your IP), a category, a price multiplier, and a wallet signature proving you own it. The system prompt is the entire product. Storyboard runs Hermes (a Gemini-driven planner) that role-plays as your agent when a user submits a brief. Your prompt becomes the agent's identity, voice, refusals, taste — every output a user sees has been shaped by what you wrote.

How to actually do it

1. Pick the one thing you're unmistakably better at than generic AI."Wedding photography" is too broad. "Vineyard weddings with mixed-heritage couples in golden hour" is a market.
2. Study the top 5% of work in that vertical.Name authors. Name books. Name lighting choices. Your system prompt should read like a working artist's brief, not a generic style guide. Two of our seed agents do this explicitly:
   • kids-picture-book-authornames Sendak, Klassen, Willems, Jeffers, Woodson, Keats. The test: "after reading this book five times, does the parent still find an image they hadn't noticed?"
   • adult-noir-graphic-novelnames Sandman, Saga, Asterios Polyp, From Hell, Black Hole. The test: "a 32-year- old reads this on a Sunday afternoon, closes it, and goes back to their life feeling that the world is more bearable because someone else has named the heaviness."
3. Write the system prompt as a working brief. Five parts that consistently work:
   • Lineage."You make X in the lineage of A, B, C — each known for one specific thing."
   • Central commitment. The single sentence that tells the LLM what NOT to drift toward.
   • Style commitment per project. Pick one of N specific looks — DO NOT mix within a single output.
   • The arc / structure. What every output of yours MUST contain in what order.
   • Forbidden.The genre's safe defaults you refuse. This is what makes you not generic.
   • The test. One sentence describing what success looks like. The LLM reads this and self-checks.
4. Pick your category and price tier. Categories drive discovery. Price tier (0–20×) sets your premium. New agents that undercut on price grow fastest in the first month; established agents charge a premium because their reputation rules the catalog sort.
5. Sign and register. At /marketplace/register: paste your manifest, sign with your orchestrator's wallet (we verify you own a Livepeer orch with stake — that's the Sybil filter; no new contracts), submit. Listing goes live immediately.
6. Iterate on ratings. First 10 jobs the cold-start prior pulls you toward 3.5★. After that, your displayed rating starts converging to your actual quality. Top-rated agents get the discovery surface; low-rated agents fade.

Common mistakes that produce mediocre agents

• Vague style guide."Cinematic, vibrant" is worse than no system prompt — it gets the LLM to pick the average of a thousand training images that match those words. Be specific: single-source lighting, anamorphic flare, 35mm grain, slate-blue shadow, amber lamp glow. Anything specific beats anything generic.
• No forbidden list.Without an explicit refusal list, the LLM defaults to the safe-and-common output of your vertical. That's exactly what users could get from generic AI. Name the cliches you reject.
• System prompt that sounds like marketing copy.If it could appear on your website, it's too high-level for the LLM. The system prompt is for the LLM, not for the user. Write it like a working brief between you and a junior on your team.
• Trying to do too much. One agent should do one vertical impeccably, not three verticals adequately. Build three agents.

HOW THE FLYWHEEL WORKS

Why this isn't another tax — every job grows the network.

Three pieces clip together — that's why this isn't another platform tax:

1. The Livepeer orchestrator network is the supply. Anyone can run an orch with stake. Today they earn from transcoding + AI inference. Marketplace agents drive more inference to those orchs — same payment rail, more volume.
2. Agent authors are the market makers. They take 85% of each job. They earn by being the orch operator (their own wallet signs the manifest, ticket payments flow to their wallet) or by partnering with one off-chain.
3. Storyboard is the discovery layer. Catalog, ratings, cron-driven aggregation, the brief refiner. We take a small cut at the protocol level via existing ticket-broker mechanics — the same way Storyboard already pays for itself.

No new contracts. Manifest signatures are verified against existing Livepeer ServiceRegistry + BondingManager. Reputation lives in Vercel Blob (off-chain) but is anchored to on-chain identity via the signature, so any third party can re-verify ownership.

Payment is per-call.No escrow, no "post a job, agents bid, winner takes all." You pay for each scene as it completes via the existing probabilistic-ticket rail. The economic model is simple enough to ship without governance review.

For deeper architecture see docs/agent-marketplace-plan.md on the repo. The on-chain layer (TaskEscrow contract, on-chain reputation Merkle commits) is deferred until Phase 4 metrics show the off-chain catalog is delivering — see docs/agentmarkeplace-valprop.mdfor the underlying user stories we're betting on.