Can non-engineers direct AI software development with Arcadia?
In brief
You do not need to be a professional engineer to understand the problem Arcadia addresses: deciding what AI should build next and whether the result works. Using the current development setup is a separate question. Cold-start setup for someone who is not the author is in progress. Using Arcadia still assumes an AI coding agent already runs against a real project you control.
- Define what the product should do.
- Test observable outcomes, not convincing demos.
- Know when technical help is needed.
The problem: directing software without writing every line
AI development for non-engineers still requires product judgment. A convincing booking-form demo does not establish that two people cannot reserve the same slot. You need to describe the intended behavior, choose priorities and recognize when verification requires expertise.
The challenge is keeping those decisions intact as the software and the agent’s context change. Product direction and technical implementation are different responsibilities.
What you may have tried
- More detailed prompts: useful for explaining intent, but the next conversation may lose the agreement.
- Demos and screenshots: show the interface; they cannot prove conflict handling, security or recovery.
- Handing everything to an expert: valuable for technical work, but your priorities and acceptance decisions still need a durable home.
Start with one observable outcome and seek technical help for the checks you cannot assess yourself.
How Arcadia connects product judgment to evidence
An Arcadia Project holds its Mission—the reason it exists—and desired Outcome. A Milestone marks progress; a Plan organizes bounded Actions with acceptance criteria. This connects a business need to work an agent can actually verify.
For a booking product, “make appointments dependable” can become an Outcome of preventing duplicate reservations. A Plan groups the form and conflict-checking Actions. You define what success looks like; evidence against each Action shows whether the delivered behavior meets it.
The benefit is continuity of direction, not automatic technical expertise. Cold-start setup for someone who is not the author is in progress. Using Arcadia still assumes an AI coding agent already runs against a real project you control. A finished no-code onboarding experience is not established. Try describing one behavior and the evidence that would convince you.
Sources and further reading
- GitHub: spec-driven development with AI — making product intent explicit before implementation.
Current capabilities and supporting evidence
- Install on someone else's machineIn progress
Cold-start setup for a person who is not the author. Early access waits on this.
Supporting evidence - Agent session managementWorking
Isolated worktrees per action, across Codex, Claude Code and OpenCode.
Supporting evidence - Completion evidenceWorking
Acceptance criteria are checked before an action can be marked done.
Supporting evidence
Capability snapshot: . See what changed →