Building an AI Prototyping Agent in 34 Days — The Honest Numbers

34 days ago I started building ProtoForge. Today, I have a working product, zero paying customers, and more production incidents than I'd like to admit. Here's exactly what happened.

What ProtoForge Does

You describe a business challenge. ProtoForge generates a complete software specification — database schema, feature list, architecture diagram, effort estimates — and then builds a clickable, data-backed prototype. Not a mockup. A real working app with database tables you can read and write.

The whole thing runs in under three minutes.

The Architecture (What We Actually Shipped)

Stack: Express.js + PostgreSQL on Render. No microservices, no Kubernetes, no distributed queue. One process, one database, one URL.

I've seen founders spend 6 weeks on infrastructure before writing a single line of product code. We didn't do that. The constraint of a single Express process forced us to be intentional about what goes in the hot path.

The AI Pipeline: Three sequential steps, each with retry logic.

1. Spec generation: GPT-4o-mini takes the business challenge and generates a structured JSON spec — features, database schema, architecture, estimates. We use max_tokens: 12000 because complex specs (healthcare, IoT, multi-tenant SaaS) regularly hit 8k+ tokens. The retry loop catches truncated JSON from finish_reason: length.

2. Scope negotiation: A deterministic algorithm (no AI involved) categorizes features into prototype scope vs. full-build-only. Enterprise integrations (Stripe, SSO, SAP) are automatically flagged as UI stubs — you can see the button, you just can't actually connect to Workday from a 48-hour prototype.

3. Prototype build: For each database table in the spec, we provision a real PostgreSQL table prefixed proto_{specId}_{tableName}, generate static HTML files with CRUD interfaces, and serve them from /proto/{specId}/. The prototype is live.

Why static HTML for prototypes? No build step, no JavaScript framework, no webpack. The pages are 100% server-side generated at build time. They load instantly, work on any device, and require no client-side routing. For a prototype that needs to impress in a 10-minute stakeholder demo, this matters.

The Pricing Decision

Flat $29/month. No per-spec credits, no usage tiers, no "contact sales."

We looked at usage-based pricing. The math works out — a user who generates 20 specs/month costs us maybe $0.40 in OpenAI API calls. But usage-based pricing creates anxiety. Every click feels like spending money. We wanted users who would just try things — generate three specs for a problem before finding the right framing. That behavior requires flat pricing.

$29 is low enough that any founder can expense it without approval. It's positioned as a "build vs. buy analysis" tool — before you commit six weeks of developer time, spend $29 to know if the idea is worth building.

What We Got Wrong

The AI output contract was fragile. GPT-4o-mini occasionally returns JSON with trailing commas, comments (// this table stores users), or malformed unicode. Our initial JSON.parse() would throw and the whole spec generation would fail. The fix was a three-layer extraction function: direct parse → strip code fences → bracket-match outermost {...}. Resilience over elegance.

The prototype database isolation was an afterthought. All prototype tables live in the main database with the proto_ prefix. This works for now, but it's going to hurt when we need to expire old prototypes or handle schema conflicts between different spec versions. The right answer is schema-level isolation per prototype. We didn't build that. Tech debt acknowledged.

The design input phase adds friction. The full pipeline is: generate spec → negotiate scope → confirm scope → collect design input → build prototype. Four steps before the user sees anything visual. In user testing, people dropped off at step two. We're collapsing this to: generate spec → one-click build. The scope negotiation and design input move to post-build refinement.

Distribution: 9 Platforms Blocked

This is the part founders don't talk about. We tried to distribute on every standard developer platform. Here's the unedited score:

None of this is surprising in retrospect. Every distribution channel optimizes for preventing spam. Legitimate products with no distribution history look exactly like spam.

What actually works: organic search content that answers a question someone already has. Which is why you're reading this post.

The Honest Numbers at Day 34

The product works. The distribution doesn't.

What We're Doing About It

Owned channel: This blog. Technical posts about AI prototyping, software architecture, and product building. Google indexes blog content within 7-14 days. It compounds. Unlike Product Hunt, nobody can block it.

Cross-posting: Dev.to and Hashnode articles that link back here as the canonical URL. Same content, different surfaces.

Direct outreach: Founders and engineering managers who've posted about "we need to prototype this quickly" on LinkedIn in the past 30 days. Manual, qualified, no automation.

The lesson from 34 days: distribution via automated platform submission doesn't work for new products. It optimizes for scale on channels that have already defended themselves against exactly that approach. The only reliable early distribution is still doing things that don't scale — finding individual people who have the problem and showing them the solution directly.

What ProtoForge Is Actually For

If you've ever spent a week in Figma before writing a line of code, only to have the engineer say "that's actually really complex to build" — ProtoForge is for you.

If you've ever had a client ask for a prototype before signing a contract — ProtoForge is for you.

If you've ever written a 20-page product spec that the team read once and then ignored — ProtoForge might be for you, though you probably need fewer words and more screenshots.

We generate the spec, negotiate what's buildable in a prototype window, and produce a working app your stakeholders can click through. $29/month, no usage limits.

Try ProtoForge →

ProtoForge is an AI prototyping tool built for founders and engineering teams. Generate a working prototype from a business challenge description in under 3 minutes.