Skip The Pitch

## Skip The Pitch

Find local businesses with weak websites — and let Codex build the replacement, end to end.

## 💡 Inspiration

Every freelancer and small agency knows the same painful truth: the hardest part of web work isn't the code — it's the sales. You spend hours cold-emailing a plumber, a dentist, a café owner, trying to convince them their 2011 website is costing them customers. Most never reply. The ones who do want to "see something first."

So we flipped the funnel. What if you showed up with the finished website already built? No pitch deck, no "trust me," no discovery call — just a real, live, better version of their site, ready to hand over.

That only becomes possible when an AI agent can do the entire build autonomously. OpenAI Build Week was the perfect forcing function: could Codex take a business it had never seen and ship a production-quality website with no human in the loop?

## 🚀 What it does

Skip The Pitch runs a full lead-generation-to-delivery pipeline in a desktop app:

Discover — Finds local businesses in any area using free map data (OpenStreetMap) and Google, filtered to real, contactable prospects.
Grade — Renders each business's existing website and critically judges it, returning a verdict: good, poor, broken, no_site, or defunct.
Rank — Prioritizes prospects by a Pitch score — a verdict-aware opportunity signal, not a vanity quality score. A broken site for a busy business beats a mediocre site for a dormant one.
Build — Dispatches Codex to research the business, design a bespoke identity, and build a complete, structurally-unique website — watched live in a Build Studio with a real-time preview stage.
Close — Tracks everything in a built-in CRM (deals, status, revenue, notes), so you go from cold lead to signed client without ever writing a proposal.

The build isn't a template with the logo swapped. Around Codex, a design-director agent turns verified business evidence into a distinct visual territory, a design-critic agent independently reviews the result for originality and conversion, and an image-curator agent selects real imagery — all before the site is handed off.

## 🛠️ How we built it

The core idea is multi-agent orchestration on the user's own machine. Skip The Pitch is local-first: builds run on your hardware using your own Codex credentials, so there's no shared build farm and no per-build cloud cost.

The stack:

Desktop shell — Tauri v2 (Rust core + React/TypeScript UI). Rust supervises the spawned Codex build process, streams its output into the Live Stage, and handles the native OS boundary.
Engine — A local SQLite database and a JSON CLI that owns discovery, website evidence, scoring, deduplication, and CRM state.
Build agents — The build is decomposed across specialized sub-agents (design direction → image curation → code generation → design critique). Codex does the actual site generation; the surrounding agents keep it honest and on-brand.
Ranking — Prospects are ordered by opportunity. Conceptually:

$$ \text{PitchScore} = w_{v}\cdot \text{VerdictGap} \;+\; w_{s}\cdot \text{BusinessSignal} \;-\; w_{q}\cdot \text{SiteQuality} $$

A worse existing site ((\text{SiteQuality}\downarrow)) and a stronger, more active business ((\text{BusinessSignal}\uparrow)) both push a lead to the top. Raw site quality saturates quickly and is only ever a junk filter — never the primary rank — because a great-looking site is a bad sales opportunity.

## 🧗 Challenges we ran into

Making agents produce distinct sites, not slop. Early builds all converged on the same "category recipe" (dentist = blue + tooth icon). Fixing this meant adding an adversarial design-critic pass and a design-director that works from business evidence rather than category defaults.
Cross-platform native processes. Supervising a spawned Codex build reliably behaves very differently in a packaged desktop app than in dev — and differently again on Windows vs. macOS. Getting the installer, runtime detection, and process boundary to fail actionably (clear errors, not silent white screens) was a real grind.
Build agents — The build is decomposed across specialized sub-agents (design direction → image curation → code generation → design critique). Codex does the actual site generation; the surrounding agents keep it honest and on-brand.
Ranking — Prospects are ordered by opportunity. Conceptually:

$$ \text{PitchScore} = w_{v}\cdot \text{VerdictGap} \;+\; w_{s}\cdot \text{BusinessSignal} \;-\; w_{q}\cdot \text{SiteQuality} $$

## 🧗 Challenges we ran into

Making agents produce distinct sites, not slop. Early builds all converged on the same "category recipe" (dentist = blue + tooth icon). Fixing this meant adding an adversarial design-critic pass and a design-director that works from business evidence rather than category defaults.
Cross-platform native processes. Supervising a spawned Codex build reliably behaves very differently in a packaged desktop app than in dev — and differently again on Windows vs. macOS. Getting the installer, runtime detection, and process boundary to fail actionably (clear errors, not silent white screens) was a real grind.
Trusting an autonomous build. An agent shipping a website with zero human review is only acceptable if you can see it working. The Live Stage — streaming Codex's real output and a live iframe of the emerging site — was as much about trust as UX.
Grading the open web. Real business websites are broken in a thousand ways (dead domains, JS-only pages, hard blocks). We built a tiered read strategy (browser → honest fetch → fallback crawl) so grading stays accurate on hostile pages.

## 📚 What we learned

Decomposition beats a single mega-prompt. Codex asked to "build a great website" produces average work. A pipeline of small agents — each with a verifiable output and an independent critic — produces work you'd actually hand to a client.
The demo is the pitch. Showing a finished artifact collapses the entire sales cycle. The product's whole thesis — skip the pitch — turned out to be the right design principle for the agent architecture too: don't describe the outcome, produce it.
Autonomy needs observability. Users don't trust a black box that spends their compute. Streaming the build made an autonomous agent feel like a collaborator.

## 🔭 What's next

Richer per-industry design territories and broader Codex tooling.
One-click handoff of the generated repo to the client.
Deeper CRM automation — from "site built" to "invoice sent" without leaving the app.

Skip the pitch. Ship the site.