48-hour prototype: a demo built with a coding agent
Two days to answer whether AI can do a job for a client: build one flow that works end to end and let them click it.
The problem
A client asks the oldest question in this job: can AI actually do this for us? A slide deck will not answer it and a month-long build is too slow to find out. You have two days to put something real in front of them.
Scope is what sinks it. What you want is one flow that works end to end, on the client's own example, that they can click through in the room. What it cannot do yet counts for less than the one thing it does being real.
Architecture
A prototype is a funnel: take the vague ask, cut it to a single flow, and let a coding agent do the typing while you steer. The second box is the one that decides it: most failed prototypes die because the scope was never cut.
The coding agent (Claude Code or Codex) is a force multiplier, not the plan. You decide the one flow and the demo; it writes and edits the code fast enough to fit two days.
What it draws on
Everything here comes from this stage of the roadmap; the project is where those courses meet.
- Claude Code: working alongside an agent
- OpenAI Codex: agents in scripts and review
What done looks like
| Requirement | Done when |
|---|---|
| Real data | It runs on at least ten real or realistic sample emails |
| Correctable | Every extracted field can be edited before saving |
| Honest | A visible note lists what is faked |
| Fast | Built within 48 hours, with the agent's plan kept in the repository |
Where to start
Cut first, build second. Write the one sentence the demo has to prove, delete every feature that is not on that path, and only then open the coding agent. Keep a running demo from hour one, even if it fakes half the work; a prototype that runs badly beats one that is nearly finished.
Every expert started right here.