My agent kept saying "done" on broken code, so I built a protocol...
I'm an iOS dev (Swift, mostly SDK and app work). Like a lot of you I started leaning on AI agents to write code - and hit the same wall every time: the agent confidently says "done, tests pass" when they don't, quietly refactors code I didn't ask it to touch, or invents an API that doesn't exist.
I got tired of that and built PayneSDD - a free operating protocol for coding agents (Claude Code first, but it pastes into any agent). It's one file of rules the agent follows. This isn't a prompt I threw together over two evenings: it's a 7-step cycle (Steps 0–6), it's been through 18 releases of iteration, and it develops itself under its own protocol - every change ships through the full cycle plus an independent review before it merges.
How it actually works:
- You write requests in plain words, like always. The protocol asks the clarifying questions, maps the open decisions, and shows you ONE plan — zero code until you say "go".
- Tasks are tiered: a typo just gets done; auth, billing, migrations, anything public-facing — all forced through the full ceremony. You don't pay full process for a one-liner.
- "Done" is the machine's word. The agent has to run your real tests/build, and an optional Stop-hook physically blocks it from saying "finished" while tests are red (3 blocks, then it releases with an explicit UNVERIFIED warning — an honest release, not a fake lock). No hook = honor-system, and the README says so plainly.
- Then an independent second agent attacks the result with a "break it" brief. Every finding has to cite a real code line or test, or it's rejected - no vibes-driven review.
- Verdict is always explicit: PASS / ITERATE / ESCALATE, plus a Done / Remaining / Open-questions checklist.
Install is one pasted message - and the first task the protocol runs on is its own installation (it interviews you about the setup, then touches your config only after your "go").
Repo: https://github.com/vlr-code/PayneSDD
I'm mostly sharing it in case it saves someone else the same headache it saved me - if it helps your dev workflow, that's a win.