Build an MVP with AI: a reproducible handoff walkthrough

Know what the example is

Reddit to AI is a public set of planning documents for a Chrome extension. The planned product captures a Reddit thread, lets the user filter and review it, and prepares the result for an AI chat. The repository states that the product existed before the workflow documents. The documents were reconstructed to show what the workflow produces.

This makes it a teaching example, not a customer case study. You can verify the files and judge whether they agree. You cannot use them to claim a development time, install count, revenue result, or improvement over another process. Those outcomes would need separate records.

Read the handoff in dependency order

Read from the user problem toward the build instructions. If you start with AGENTS.md, its rules can look sensible even when they do not match the product. The earlier documents explain why the later rules exist.

Stop when a dependency is weak. If the research does not support the user problem, polishing the technical design only makes the wrong idea more detailed. If the PRD still has open scope decisions, the coding agent should not resolve them silently.

Run a consistency review

Read across the files for contradictions. In the example, the PRD requires a preview before content is sent. The technical design must therefore include a preview surface, and the agent instructions must not permit a shortcut that sends content directly. One decision should survive the whole handoff.

• Every must-have feature has an implementation path and at least one observable check.
• Every component in the technical design supports a requirement or a stated operational need.
• Privacy and security promises match the chosen data flow.
• The agent commands, file paths, and test commands are real for the target project.
• Open questions remain visible. They are not rewritten as decisions.

A long document can still fail this review. Completeness is not a word count. It is agreement between the problem, scope, design, and evidence.

Give the agent a bounded first task

Do not finish the review with “build the whole MVP.” Ask the agent to read the handoff, summarize the constraints, and propose the first small phase. The phase should name the files it expects to change, the behavior it will produce, and the commands that will check it.

Read AGENTS.md, the PRD, and the technical design.
Summarize the first build phase and its acceptance checks.
Do not edit files until I approve the phase.
After approval, implement only that phase and report the checks you ran.

This prompt does not make the agent correct. It gives you a review point before changes begin and a clear boundary when the task starts to grow.

Record evidence, not confidence

At the end of the phase, keep the changed-file list, test output, browser or device checks, and unresolved risks. “The agent says it works” is not evidence. A passing command is useful, but it also has a limit: a typecheck cannot prove that the interface is understandable.

Use the product through the same path a user will take. Deliberately test one failure. For the example extension, that means checking the normal scrape and preview flow, then breaking the automatic paste path to confirm that the copy fallback remains usable.

Save that result in the project handoff before starting the next phase. The annotated PRD guide explains how to make those checks trace back to product requirements.