Codebase Exploration with AI

Welcome back! We are going to investigate one small feature: deleting a feed entry. We are not implementing it. We only want to know where the change belongs, what it could break, and what questions must be answered first.

The feature is intentionally bounded. It covers deleting one entry, preserving the rest of the feed, and cleaning up a photo when necessary. Bulk deletion, undo, synchronization, and unrelated modernization stay out of scope. I’m going to write a prompt for this.

You can find the delete-feed-entry-feature-brief.md file in the AI Resources section of the student materials for this module.

Using BabyTrack and the delete-feed-entry feature brief, create a source-backed feature-readiness plan. Identify change points, risks, testing needs, and open questions with file citations. Do not implement the feature. Return the complete Markdown for feature-readiness-plan.md.

Awesome. Let’s take a look at what it returned, and we now have our delete feed entry feature readiness plan, and it looks pretty robust.

I do not need the AI to choose a particular implementation. I need it to identify the relevant boundaries, support its claims, and admit what remains unresolved.

The next step is having the AI recheck the plan, flag any unsupported claims, missed change points, unnecessary scope or risks, or any questions, and to cite the source. So I’ll go ahead and do that.

Re-check the plan against BabyTrack and the feature brief. Flag unsupported claims, missed change points, unnecessary scope, risks, and open questions. Cite the source.

Fantastic. Now, you could have used a fresh conversation because it could give you a more independent review. But nevertheless, the reviewer still does not make the decisions. I verify the claims and choose what belongs in the final plan.

So let’s say I agree with this plan. I’m going to write out another prompt to finalize it, save it to a markdown file, and to preserve any unresolved questions or scope that we may have rejected.

Revise `feature-readiness-plan.md` using only the review findings I accepted. Preserve unresolved questions and rejected scope. Return the complete final Markdown without implementation code.

We now have one reviewed readiness plan without extra draft files or a long chain of prompts.

You traced a feature through a mixed codebase, reviewed the risks independently, and kept the final decisions with a human. The result is ready for detailed development planning, not automatic implementation.