Case study · the platform · 2026-04-20

Escape 9 to 5 built itself in 48 hours.

Every module, every page, every commit — from an autonomous build loop running on a task queue.

Tasks shipped autonomously
48+
setup, rewrites, polish, modules — all hands-off
Cron fires per hour
2
at :17 and :47 — off-beat to avoid rate-limit convergence
Human touch
~3%
direction + corrections in Telegram; build itself was hands-off
Time from zero to 12-module platform
48 h
2 calendar days, mostly while Loya worked on other projects
tl;dr

This entire platform was shipped by an autonomous Claude Code agent reading a BUILD_STATE.md task queue, picking the next task every 30 minutes, verifying with tsc + build, committing, and pinging Loya on Telegram. 48+ tasks shipped across 48 hours. Zero direct code-by-hand.

The problem

A one-person founder can't hand-code a full membership platform while also running their main product. But you can't hire it out either — the product is specific, the voice is specific, the stack is opinionated.

The setup

Two files, one cron, one agent.

BUILD_PLAN.md — written once. 100 lines. Defines the goal, stack, constraints, non-goals. The agent reads this every fire but never edits it.

BUILD_STATE.md — the live task queue. ## In Progress, ## Next Up, ## Completed, ## Blockers, ## Consecutive Failures Counter. Agent edits this after every commit.

A cron firing every 30 minutes (at :17 and :47 — off-beat to avoid rate-limit convergence with everyone else on :00 and :30). Wakes the agent with a fixed prompt: read state, pick next task, execute, verify with tsc + build, commit, update state, ping Loya.

The auto-brakes

  • PAUSED: true flag in BUILD_STATE → agent silent NO-OP
  • Consecutive Failures Counter ≥ 3 → agent auto-pauses and pings for help
  • Queue empty → agent pings once, then NO-OPs until queue grows

These brakes mean Loya could go to sleep for 8 hours and wake up to 10 more tasks shipped — or to a single polite "blocked on X, what do you want to do" message.

What Loya did

Course-correct in Telegram. "This reads like an ad — make it less polished" → agent rewrote 5 modules in the next hour. "Seedance caps at 15s, not 20s" → agent patched the module + re-committed in 2 minutes. The corrections got faster as the agent's memory accumulated (feedback memories in its vault).

The result

This whole site — landing, every module, admin panel, community, Telegram linking, Stripe checkout, R2 skill-install pipeline, OG images, changelog, founding-member coupon infrastructure — shipped while Loya worked on a different product (rewriting the LYRC editor).

What modules taught it

  • Module 12 (Autonomous Build Loop) — the framework: plan + state + cron prompt + auto-brakes
  • Module 08 (Autonomous Agent on Telegram) — the agent sitting on Telegram so Loya could course-correct from anywhere

Install both, write your own BUILD_PLAN + BUILD_STATE, and you can do this for any project where the punch list is executional — an app, a course, a content series, an agency delivery pipeline.

modules that built this