Before an AI agent builds, reconcile the specs and plan checkpoints

Devin introduces a smart personal trainer app being developed from specifications. Before handing three design documents to an agent to build, he wants to resolve missing details and contradictions.

He gives Gemini references to all three documents and explicitly asks it not to create anything, build or write code. The agent lists information still needed and inconsistencies across the specifications.

For each issue, Devin asks what the agent would have assumed if told to build immediately and what it recommends instead. Comparing those answers helps him identify choices that otherwise might have been made silently.

He considers either editing the three documents or adding a fourth document with assumptions. Gemini recommends updating each source document in place; Devin reports its reasoning that details kept together are less likely to be missed.

To preserve the existing work, Devin asks Gemini to back up the documents, propose an implementation plan for the edits, make the changes and review them. After he is satisfied, he deletes the backups; he mentions Git as another way to preserve versions.

In a new session in the same workspace, he supplies the revised specifications and again asks for review, not code. He asks whether to combine the documents and whether a separate app implementation plan would help.

Gemini recommends keeping the three specifications separate and organized by subject, and adding a distinct implementation plan. Devin explains this as a way to make relevant information easier for the agent to find.

Devin asks whether the agent should build the entire application in one pass or proceed in steps. The recommended approach uses phases with verification between them, so a problem is easier to isolate before the next phase begins.

The proposed gates begin with shared types and the database, followed by a human check of the types and database seed. The next stage covers server APIs and the AI coach, then state infrastructure and workout execution, and finally interface assembly and systematic tests.

Devin asks for a human checkpoint at each stage. He also describes Antigravity’s ability to run the app in a browser, test functionality and attempt fixes, while noting that he may still click through it himself.

After the agent writes the plan, the workspace has four documents: the three design specifications and the implementation plan. Devin then asks how to return to a working state if a later phase breaks something.

He asks the agent to add version control and rollback checkpoints to the implementation plan. With the planning documents updated, the application can proceed to the build stage.