Codex Goal Mode Took My Funnel App to Launch-Ready

FunnelFlow is a side project from February: a drag-and-drop marketing funnel mapper built on React Flow and Firebase, positioned as a free alternative to Funnelytics. A collaborator and I got it to a working canvas editor, then life moved on and the repo sat idle for four months.

This week I revived it with a different experiment. Instead of pair programming feature by feature, I gave Codex CLI's goal mode one instruction: get this app ready for a public launch. Then I mostly stayed out of the way.

What Goal Mode Changes

Normal AI coding is conversational. You ask for a change, review it, ask for the next one. You are the loop.

Goal mode inverts that. You hand the agent a destination instead of a task list, and it plans the milestones, does the work, checks its own output, and keeps going until the goal is met or it needs a decision only you can make. The unit of delegation stops being "write this function" and becomes "close the gap between this repo and a launchable product."

That framing matters because launch-readiness is exactly the kind of work humans procrastinate on. It is not one feature. It is forty unglamorous things: legal pages, SEO metadata, analytics wiring, security rules, empty states, error retries. Perfect work for an agent that does not get bored.

What It Shipped

The working tree now holds roughly 20,000 changed lines. The highlights:

• A public marketing landing page, blog with three SEO articles, privacy policy, and terms of service — all as lazy-loaded routes so the editor bundle stays lean
• Read-only funnel share links, private by default, enforced with Firestore security rules rather than UI politeness
• Shared funnels can be marked as templates that visitors duplicate into their own account, with the intent surviving the sign-in redirect
• A first-run onboarding checklist plus a sample marketing funnel so new users do not land on a blank canvas
• An analytics bridge for GTM or Plausible that sanitizes event properties so user-entered funnel data never leaks into tracking
• Prerendered static HTML per route, sitemap, RSS feed, robots, Open Graph tags, and JSON-LD
• The test suite grew from 2 test files to 28, covering routes, hooks, sharing, storage fallbacks, and SEO utilities

It also deleted things: a dead screenshot microservice, nine stale build scripts, three orphaned utility modules. Agents that only add code are easy to find. Agents that reduce surface area are rarer and more valuable.

The Launch Gate Is the Interesting Part

The piece I keep thinking about is not any feature. It is that the agent wrote its own definition of done: fifteen verification scripts wired into a single npm run verify:launch gate that checks lint, tests, Firestore rule invariants, bundle budgets, SEO output, route artifacts, blog content quality, Docker config, and analytics event wiring.

An autonomous agent is only as trustworthy as the checks it leaves behind. The deliverable is not the code. It is the code plus the harness that proves the code.

I did not take its word for it. I ran the full gate myself on a fresh terminal: every check passes, zero dependency vulnerabilities. That is the same trust-but-verify habit I use everywhere with AI output — the agent's claim is a lead, the passing run on my machine is the finding.

What Still Needs a Human

Goal mode did not launch the product, and being honest about that line matters. Still on me: creating the real Firebase project, the Google OAuth consent screen, App Check, the domain, and pointing analytics at a real destination. And the code itself is uncommitted until my collaborator and I review it — an agent that pushes to a shared repo without review is a bug, not a feature.

Where FunnelFlow Goes: Funnels as GTM Analysis

Shipping the app is step one. The product idea I actually care about is making funnel maps do go-to-market analysis instead of just documentation.

Most funnel tools live after the fact — attribution dashboards that tell you what already happened. The gap is before the spend: a place where reach, conversion assumptions, and budget live on the map itself, so the plan is an explicit model you can interrogate rather than a slide.

The roadmap that follows from that framing:

• Assumptions first. Every edge carries an estimated conversion rate; every source carries reach and spend. The map computes expected volume at each stage before a dollar moves.
• Plan-versus-actual calibration. Import actuals from GA4 or a CRM and diff them against the assumptions. I learned this one building our baby sleep tracker: when actuals miss the plan consistently in one direction, the plan is miscalibrated, not the team. The tool should say that and offer the corrected number.
• Scenario comparison. Duplicate a funnel, change the assumptions, and compare expected outcomes side by side — more traffic versus stronger offer versus fewer steps.
• AI critique. An agent reads the map and flags what would make the plan wrong: missing tracking events, conversion rates far outside channel norms, stages with no owner.

That turns a diagramming toy into a cheap GTM analyst — which is the version worth marketing.

The Takeaway

The bar for resurrecting a shelved side project just dropped again. Four months of "I should really finish that" became a weekend of agent time plus an afternoon of my review. The skill that mattered was not prompting; it was defining the goal crisply, insisting on a verification gate, and knowing which decisions to keep human.

I write up experiments like this, including the ones that fail, in the free AI community. No course, no paywall. Just people learning by shipping useful things.