Cloudflare shipped EmDash in April 2026 — an open-source CMS written in TypeScript, built in ~2 months by AI coding agents. It's a genuinely impressive achievement and a real signal of where the industry is going.
But it also surfaces a question that the AI coding conversation has been avoiding: what happens after the AI ships the first version?
The "plans that read well don't build well" problem
There's a failure mode I keep seeing in AI-assisted codebases. The initial build is fast. The prose in the plan reads authoritatively. The code compiles and the tests pass. Three weeks later, the second engineer tries to extend it, and nothing quite fits — because the agent's narrative was persuasive without being correct about the underlying constraints.
This isn't a model problem. Frontier models will keep getting better at writing plausible code. It's a workflow problem. The missing layer is the one that turns ephemeral agent sessions into durable, reviewable architectural decisions.
What the missing layer looks like
We have been building rp1 with this exact gap in mind. Three ideas, each directly addressing a specific failure mode.
1. Constitutional prompting
Most "prompt engineering" is additive — you stack instructions on top of a model and hope. Constitutional prompting is subtractive: workflows encode the patterns an expert would follow as constraints. /build isn't a prompt, it's a pipeline:
- Generate a blueprint from requirements
- Form a hypothesis about the existing codebase
- Validate the hypothesis against actual code before writing anything
- Implement against the validated plan
- Run verification
The hypothesis validation step is the one that catches the "plan reads well but is wrong about your ListView" class of bug.
2. Knowledge-aware agents
Most AI coding sessions start blank. You re-explain your architecture everytime. rp1's /knowledge-build runs once and maps your codebase into a persistent knowledge base that every subsequent command inherits.
The practical effect: you stop getting generic advice that ignores your patterns. Every /build starts with full awareness of the actual system, not an imagined one.
3. Durable artefacts
Every rp1 workflow produces inspectable design documents — requirements, design, hypothesis, verification, reports — attached to the project, not trapped in chat scrollback.
This is the onboarding primitive. When the second engineer joins an
AI-built codebase, they can read what was decided and why instead of re-prompting their way to an understanding.
Try it
rp1 is open source and works across Claude Code, OpenCode, Codex, and GitHub Copilot CLI. Same workflows, different harnesses.
The full write-up on how this plays out specifically for EmDash-style codebases is on our blog: rp1 on EmDash — the workflow layer that makes AI-built codebases navigable.
If you're maintaining a codebase an agent wrote, Prem and I would genuinely like to hear what's broken. That's the feedback that's shaped everything we've built so far.
rp1 is built by Prem Pillai (@cloud_on_prem) and Mahesh Shivamallappa (@maheshs786).
Top comments (0)