9.4 billion tokens. 66 hours. 114 tickets.
Is that still "vibe coding"?
Yesterday, we finished a full streaming platform build. A UGC Netflix for AI films. Uploads, streaming, comments, creator payouts, admin moderation, copyright disputes. Not a demo, a real product.
I didn't build it ticket by ticket. I set up an autonomous loop in Claude Code, gave it the spec, and let it run. I made the calls at the decision points. It did the work.
The run:
- 114 tickets. Built, reviewed, tested, shipped.
- 66 hours of autonomous loop build time.
- Claude Code on two models: Opus 4.8 and Fable 5.
- 127 agents, each with one job.
- 1,200+ background jobs across the run.
- 9.4 billion tokens processed.
- 3,900+ automated tests, all green.
Every ticket ran the same loop: build, review, fix, verify, ship. Not build and hope. Reviewed by adversarial agents whose only job is to break the work before a user can.
One example. A copyright-takedown feature passed 3,881 tests. Green across the board. The review still found a way for an admin to republish content that was subject to a court-ordered permanent takedown. A legal landmine the tests never saw. Caught and fixed before it shipped.
That is the difference between vibes and engineering.
This held up because of what came first:
A brief.
A scope.
User journeys.
A 100+ features backlog.
A full PRD. Almost 170 pages.
The loop was fast because the spec left nothing to guess.
And the spec came from experience, not vibes:
12,000+ hours building apps myself.
170 products from other developers, studied.
30+ we shipped ourselves.
So call it what you want.
From where I sit, this is the speed of AI and the judgment of experience, in one build.
That's the whole game.
Next up: the QA pass. An agent walks the live app ticket by ticket against the acceptance criteria, the way a human tester would. Click by click.
I'll post what it finds.
Now, your turn.
Ask me anything:
- How the loop is set up.
- What actually goes in the spec.
- Where AI still needs a human.
- Anything else...
From our 114-ticket AI build: one ticket scored 90/100 in review and still carried an exploitable phishing path.
New rule: pass the bar AND zero open majors, or it doesn't ship.
Averages hide landmines.
Unexpectedly, Anthropic reset my weekly Fable 5 usage after I exhausted it yesterday.
Iโve got a few tasks left before the 19 July cut-off:
- Memory audit
- Refine my voice profile
- Adversarial reviews for two apps
- Skills review and optimisation
- Distribution strategy for https://t.co/98dzMZ0b7o
What are you using Fable 5 for before sunset?