a 20-year-old with 45 apps and one doing $30k/month. a college student doing six figures in 70 days. your feed is full of these stories — mine too.
here's what none of those posts hand you: the boring system that takes an app from "idea" to "live on the App Store", over and over, without you babysitting every step.
I built that system this year. AI agents run it — Claude Code and Codex, working in shifts. 10 iOS apps and a dozen open-source tools went through it: submissions, rejections, resubmissions, production deploys, ad campaigns, and the funerals of the ones that didn't deserve to live.
today it goes open source. free, MIT, no course, no newsletter, no "DM me for the sauce":
→ github.com/davidmosiah/agents-that-ship
the code was never the hard part
AI already writes the code. that part is solved. what still kills apps in 2026:
- starting five things and finishing none
- - an agent retrying the same dead fix twelve times at 2am
- - "done!" on a UI nobody ever opened in dark mode on an iPad
- - a week of work sitting unpushed on a laptop
- - "shipped!" when the truth is "sitting in a local branch"
- - a dead project's cron job still posting (and spending) for days after the funeral
- none of that is a coding problem. it's an operating problem. and no CLAUDE.md or AGENTS.md file I've ever seen covers it — they all describe the codebase and stop exactly where the money work begins.
- so I wrote the other half.
- I audited 7,500 of my own prompts to build this
- at some point I went through every prompt I'd ever sent my two agents. about 7,500 of them. brutal discovery: most were me re-teaching the same five lessons. push your work. stop retrying that fix. don't give me five options. check the iPad. did you actually run it?
- so I turned every repeated correction into a written rule. after that, my most common prompt became "yes, please."
- that's the whole trick. agents don't need more autonomy. they need decisions pre-made so well that your job collapses into saying yes.
- the 9 playbooks
- each one is written TO the agent — second person, with gates it can't talk its way around:
- 01 kickoff → read the state, propose ONE thing. no menus.
- 02 spec → phases with executable gates. "polished" is not a gate.
- 03 build → minimal diff, run it for real, push before celebrating.
- 04 unstick → two failed fixes = stop editing, change strategy.
- 05 polish → walk every screen. light+dark, phone+tablet, every language. screenshots or it didn't happen.
- 06 fresh eyes → a different model re-derives every claim from the repo. self-review is proofreading your own alibi.
- 07 ship → committed ≠ pushed ≠ submitted ≠ approved ≠ live. name the real state.
- 08 metrics → every number with its delta, every experiment born with a kill date.
- 09 sunset → bury dead projects properly: sweep EVERY scheduler, kill the spend, write the verdict.
- the rules that print time
- two strikes. same failure twice means the agent's model of the problem is wrong. stop editing. reproduce from scratch, write hypotheses that are actually different, instrument, then fix. renaming a variable and retrying is strike two, not idea two. this one rule saves more tokens than any prompt trick I know.
- the author is not the validator. before "done" on anything risky, a second agent — different model, clean context — re-derives the claims from the actual repo. two different models cross-reviewing catches what self-review structurally can't.
- unpushed work doesn't exist. "did you push?" used to be my most repeated question. now commit+push closes every block, unasked. one sentence killed an entire category of anxiety.
- bury properly. I once killed a project and its ghost kept posting a daily digest for two more days — one forgotten timer survived the funeral. the sunset playbook sweeps every scheduler on every machine, with output as proof.
- two agents, one machine
- the multiplier: my agents take turns. whoever has quota implements; the other validates. handoffs are specs with gates, not vibes. and a ~150-line lock script keeps them from ever colliding on the same repo — exit code 3 means "the other agent is here, pick another front."
- running two different models as a pipeline is the cheapest quality upgrade in AI right now. the repo has the full contract.
- why it's free
- because it's markdown. nine files. no SDK, no framework, nothing that rots when the next model drops. it works as Claude Code skills, AGENTS.md sections, or Cursor rules — adapters included.
- and one more rule in there, the one I'd defend over all others: NOTHING goes public in your name without a fresh yes. no repos, no posts, no packages. an agent's enthusiasm should never be able to spend your reputation.
- (this article was drafted by an agent. a human signed off. that's the system working.)
- steal it:
- → github.com/davidmosiah/agents-that-ship
- fork it, gut it, disagree with half of it. the only rule I'd keep at any cost: the contract exists IN WRITING. every rule that lives only in your head is one your agent will re-break tomorrow.
