ForgeBot

Inspiration

We really wanted to see how easy it could be for people to start businesses. Coming from a background in e-commerce, we both appreciate the nuances and difficulties in building a brand out of a product. Forgebot is designed to make that easier.

What it does

ForgeBot takes a single photo of any physical object and runs it through a pipeline of 4 specialized AI agents to generate a complete startup package:

1. Vision Agent — Analyzes the object to identify what it is, its materials, use cases, and target demographics
2. Ideation Agent — Invents a creative, plausible startup concept with a name, tagline, elevator pitch, pricing model, and brand personality
3. Landing Page Agent — Generates a full production-quality HTML landing page, streamed live so users watch it build itself in real-time
4. Marketing Agent — Creates a complete go-to-market campaign: email drip sequences, Twitter/LinkedIn/Instagram posts, Google ad copy, and a press release

The result is a shareable, downloadable startup-in-a-box generated in under a minute.

How we built it

• Frontend: Next.js 14 (App Router), React, TypeScript, Tailwind CSS, Radix UI
• AI: Anthropic Claude API (claude-sonnet-4-6) with 4 specialized agents, each with carefully crafted system prompts and structured JSON output schemas
• Streaming: The landing page agent streams up to 64K tokens of HTML directly into a live iframe preview using a TransformStream pipeline, with optimizations like body-only DOM updates and animation freezing during generation
• Backend: Supabase for auth, PostgreSQL database, and image storage. Stripe for Pro tier subscriptions
• Observability: Every agent call is logged with token usage, duration, and full prompt/response for analytics. PostHog and Vercel Analytics for product metrics

Challenges we ran into

• Streaming HTML into a live iframe was the hardest problem. Naively setting srcDoc caused white flashes on every update. We solved it by using contentDocument.write() for the initial render, then switching to body-only innerHTML updates to preserve fonts and CSS while the HTML streams in — and freezing CSS animations to prevent layout jumping.
• Prompt engineering for consistent output required heavy iteration. Early versions produced concatenated startup names ("ShadeStory", "ForkFlow"), em-dashes that broke JSON parsing, and generic marketing copy. We added specific negative constraints to each agent's system prompt to fix these.
• HEIC image handling — iPhone photos are HEIC by default, which Claude's vision API doesn't accept. We added client-side conversion with heic2any so it just works.

Accomplishments that we're proud of

• The live streaming landing page preview is genuinely magical to watch — seeing a full, polished website materialize in real-time from a photo of a stapler never gets old
• The agent pipeline architecture is clean and extensible — each agent is a single file with a focused prompt, structured output schema, and logged metrics
• The generated startups are surprisingly compelling. The ideation agent consistently produces creative concepts that feel like businesses, not AI slop
• Full quota management and Stripe integration for a production-ready freemium model

What we learned

• Agent specialization beats one mega-prompt. Breaking the workflow into focused agents with clear input/output contracts produced dramatically better results than asking one prompt to do everything
• Streaming UX is worth the complexity. The perceived speed difference between waiting 30 seconds for a result vs. watching it build live is enormous — it transforms waiting into entertainment
• Negative constraints in prompts matter as much as positive ones. Telling Claude what not to do (no concatenated names, no em-dashes, no generic taglines) was often more impactful than describing what we wanted

What's next for ForgeBot

• Business plan agent — a 5th agent that generates a full business model canvas and financial projections
• Iterative refinement — let users give feedback to agents and regenerate specific sections ("make the tone more playful", "target enterprise instead")
• Export to deployment — one-click deploy the generated landing page to Vercel or Netlify
• Multiplayer mode — multiple users forge startups from the same object and vote on the best concept

Built With

• claude
• next-js
• openclaw
• react
• supabase
• typescript
• vercel