36 Hours of Claude Code to a Production SaaS: The "Meta-Planning" Unlock

Claude Code ran for 36 hours straight and produced a payment-ready SaaS. The hack wasn't a better mega-prompt — it was adding one meta-planning step before the master plan, where Claude Code audits what tools, MCP servers, skills, and browser access it'll need to replace every task a human would normally do (API keys, dashboards, deployments, monitoring) and wires them up first.

The Source Methodology

Two existing methods recombined with one addition:

Get Shit Done framework — structured planning and execution for AI agents
Be Mad method — aggressive autonomous task completion
The new step — pre-planning meta-audit of tools/skills/MCP access

The Breakthrough: Meta-Planning Before Master Planning

Standard flow:

[Prompt] → [Master plan] → [Execute]

The unlock adds a step:

[Prompt] → [Meta-plan: what tools do I need?] → [Wire them up] → [Master plan] → [Execute]

The meta-planning step forces Claude Code to enumerate every human-in-the-loop bottleneck before it starts coding, including:

API keys needed (Stripe, Vercel, PostHog, etc.)
External dashboards to configure
Accounts to create
Deployments to trigger
Monitoring to set up
Browser access to grant

Once the tool inventory is complete, Claude Code configures the MCP servers and browser automations that will let it do those human steps itself during development.

What Claude Code Can Now Do Autonomously

With the tools pre-wired, during the 36-hour run Claude Code could:

Navigate the Stripe dashboard to configure products, pricing, webhooks
Navigate the Vercel dashboard to configure deployments, env vars, domains
Navigate the PostHog dashboard to set up analytics and user monitoring
Test external integrations end-to-end before attempting deployment

Why This Unlocks 36-Hour Runs

Long Claude Code sessions fail most often because the agent hits a human-required blocker:

"Go create an account at X"
"Paste this API key from the dashboard"
"Click 'deploy' on Vercel"

Each blocker freezes the session. The agent waits for the human — sometimes for hours — and context degrades while it's waiting.

The meta-planning step pre-identifies every blocker and installs the tools that remove it. Now the agent hits a dashboard step and handles it via browser automation instead of waiting.

Without meta-planning	With meta-planning
Agent hits blocker → waits for human	Agent hits step → performs it via pre-wired tool
36 hours stretches into weeks	36 hours = actual wall clock time
Context degrades during waits	Continuous execution
Many small "did you finish?" check-ins	One big outcome

The Generic Recipe

For anyone building long-running Claude Code sessions:

1. Add a meta-planning phase

Prompt: "Before you write any code, enumerate every external system this project will touch. For each, identify the tools, MCP servers, browser access, or skills you'll need to complete that step without me."

2. Make Claude install those tools itself

"Configure and test each tool before you begin the master plan. Confirm you can log into each dashboard and perform a basic operation."

3. Now do the master plan

Only after the tooling audit is complete should Claude start scoping the actual build.

4. Let it run

With every human-blocker pre-removed, long autonomous runs become viable.

Caveats

Giving AI browser access to your Stripe dashboard is non-trivial trust — scope credentials, use test mode, review actions
Meta-planning is only as good as the inventory — if the agent misses a tool, the session still blocks there
36 hours ≠ free — that's a significant token bill. Estimate cost before kicking off.

Key Takeaways

The unlock for 36-hour Claude Code runs is a meta-planning step before the master plan
That step enumerates every tool, MCP server, skill, browser access Claude Code will need to replace human intervention
Claude Code autonomously configures Stripe / Vercel / PostHog dashboards during the build
The underlying principle: remove human blockers proactively instead of hitting them reactively
Scope credentials carefully — this gives AI direct dashboard access to revenue-critical systems

Resources

Claude Code — Anthropic's terminal-based agent
Model Context Protocol (MCP) — The standard Claude Code uses to connect external tools
Stripe, Vercel, PostHog — The specific services autonomously configured in the 36-hour run