How Matt Mazur Uses Claude Code to Ship Fixes to His SaaS Preceden
Built by Matt Mazur · · Source: X
Result: According to Mazur, the system shipped 1,531 PRs in about a month across 20 batched deploys. None of the deploys were reverted, and he reports no downtime, no security issues and no bug reports caused by the changes.
What they built
Matt Mazur no longer has time to work on Preceden, a SaaS timeline maker he started 17 years ago as a Ruby on Rails side project. About a month before posting, he started letting Claude Code ship changes to it on its own. He says it has merged 1,531 PRs. The biggest groups were support article improvements (437), new test coverage (392), Ruby bug fixes (162), and dead code removal and refactors (144). He deliberately kept each change small so he could work out the system around it first.
How it works
Mazur says he didn't build a custom harness. Everything runs in the Claude Code desktop app, and he keeps macOS caffeinate running so the Mac never sleeps. Each scheduled routine starts a fresh session with a written playbook:
- SEO agent (hourly): works on support docs, structured data, sitemaps and landing pages.
- Code agent (hourly): fixes bugs, using a test that fails before the fix and passes after it as proof. It also adds tests, removes dead code and fixes performance issues.
- Dependency agent (nightly): updates one package per night and handles security alerts first.
- Other agents: one picks topics for AI-generated history timeline pages, one tracks search rankings, and a weekly retro agent audits the rest.
The agents can't touch billing, auth or email logic, and they can't change the database schema except to add indexes. Each change gets its own branch and PR into a dev branch, and a security review runs before each merge. Twice each weekday, a deploy agent merges dev into master. That step runs the full test suite, a Claude security review, a second review by Codex and CI. The agent then deploys through Hatchbox, watches the logs and Sentry, and can roll back the whole batch. When the agents need a decision, they file an "ask" with a recommendation. Mazur answers these each morning, and his answers become standing rules that every agent reads first. The agents also have access to production Postgres (read-only), Search Console, Help Scout and other services.
The result
Mazur reports 20 batched deploys so far, ranging from 5 to 275 changes each, and none were reverted. He also reports no downtime, no security issues and no bug reports caused by the changes. He expects these improvements to start showing up in MRR. He calls the experience "weird and somewhat unsettling" and says Claude is better than he is at the work it has taken on.
How to build something similar
- Start with one daily scheduled routine that opens a PR making a single small improvement, and review every PR yourself.
- Write clear limits into the playbook, like areas the agent can't touch and a maximum change size, and add a kill switch before you give it more autonomy.
- Have the agent write down the questions it can't answer. Turn your answers into standing rules so it doesn't ask the same thing twice.
- Give the agent more access and responsibility only as your test coverage, reviews and rollback process prove reliable.
The original post
About a month ago I started experimenting with letting Claude Code autonomously ship changes to my SaaS product. Here's what it has shipped, how the system works, and what it has been like to watch. For context, the product is Preceden, a SaaS timeline maker. It's a 17-year-old side project and one of the first Ruby on Rails apps I ever built, so you can imagine why Claude has had plenty to keep it busy. Over the last month it has shipped 1,531 PRs covering support article improvements (437), test coverage additions (392), Ruby bug fixes (162), dead code removals and refactors (144), JavaScript bug fixes (113), marketing and landing page improvements (98), in-app UI and copy fixes (66), technical SEO fixes (45), accessibility fixes (30), performance fixes (16), new support articles (11), dependency updates (11), and miscellaneous improvements (6). These were intentionally limited in scope so I could figure out how to build the system around it before expanding to more ambitious projects. Importantly, I didn't build some fancy harness for this. It all runs in the Claude Code desktop app. I keep caffeinate (a built-in macOS command that prevents your Mac from sleeping) running, so Claude is able to do its thing 24/7. Here's an overview of the process, which Claude helpfully put together for me: - Several scheduled Claude Code routines run on my Mac around the clock, each a fresh session with a written playbook - SEO agent (hourly): support docs, structured data, sitemaps, landing pages, up to two changes per run - Code agent (hourly, offset 30 minutes): bug fixes with a failing-then-passing test as proof, test coverage, dead code, performance - Dependency agent (nightly): one gem or npm bump per night, security alerts first, only releases at least 14 days old - Listings agent (daily): measures how our AI-generated history timeline pages perform in search, mines 20 new topics from search demand, and queues them - Rank tracker (daily) and a weekly retro agent that audits the other agents and files recommendations - Guardrails: no billing, auth, or email logic, no schema changes except indexes, size caps per change, and a kill switch file that halts everything - Every change is its own branch and pull request, merged into a dev branch, never master. A mandatory security review runs before each merge, and any finding halts the system until I look - Twice each weekday (9am and 2pm, while I'm around) a deploy agent merges dev into master through a PR: full test suite, a Claude security review of the whole batch, an independent Codex review as a second opinion, then required CI. Master is protected, even from me - It then triggers the deploy through the Hatchbox API, watches logs and Sentry for 30 to 45 minutes, and can revert the whole batch. 20 batched deploys so far, from 5 to 275 changes each, none reverted - Everything is tracked in plain files: a ledger entry per run (about 1,000 runs so far), a bank of candidate ideas, read-backs to check later whether an SEO change worked, and a lessons file that gets distilled into shared rules - Reliability is mechanical where possible: atomic locks so agents don't collide, a run can't finish without writing its ledger entry, a nightly job reconciles merged PRs against the ledgers, and each agent stops a sibling that has hung - When it needs a decision it files an "ask" with context, options, and a recommendation (67 so far). One Slack message at 9am tells me how many are waiting, and I work through them one at a time in a Claude Code session that verifies each one and then executes my answer - My answers become standing directives (25 so far) that every agent reads first, so the same question doesn't get asked twice - Anything touching production data becomes a ready-to-run command or script that I run myself I've also given it API access to almost every service Preceden relies on, so it can make better decisions and operate mostly autonomously: - GitHub (gh CLI): every change is i
Want to build something like this?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week.
Join Shipping Skool — Free