Before You Change One Line of AI-Generated Code, Save These 7 Things

Before making any modifications to AI-generated code, it is crucial to take certain precautions. Your application functions correctly, but a single small change can lead to unexpected consequences. The layout may become disorganized, buttons may stop responding, and features you have recently completed might disappear. To prevent this, duplicate the error message into a new chat and request another solution.

You will then encounter multiple issues and lack a clear way to revert to the previous state. AI assists in generating changes quickly, but it does not guarantee their safety. Before initiating any modifications, dedicate five minutes to creating a concise project handoff. This does not require a specific tool; a Markdown file within your repository will suffice.

Save the current version of the code, ensuring the ability to undo the subsequent change. If utilizing Git, create a checkpoint commit that includes only the files intended for modification. Avoid including sensitive information, generated files, or unrelated alterations. If you have not used version control yet, make a backup of the project prior to editing.

Instead of merely stating the desired change as "Improve this dashboard," specify the exact change requested: "Add a date filter above the activity table, keeping the existing layout and table functionality unchanged." Provide AI with clear boundaries and objectives to facilitate easier review and testing. Describe the expected behavior clearly, such as: "When a user selects a start and end date, display only activity within that range, including both boundary dates.

If no matches are found, display an empty state, and clearing the filter should restore all activity." This enables you to test the changes against a defined standard rather than relying on subjective judgment.

Identify all the files involved in the targeted behavior, not just the file where the issue was initially observed. For instance, a filter might encompass the UI component, data-fetching logic, and existing tests. If uncertain, ask AI to review the project and determine the relevant files before making any edits. Exclude sensitive information, private user data, or production secrets from the documentation.

Record known issues separately from the change to prevent the model from expanding a minor task into an extensive cleanup project. Document what you have already attempted when an approach fails, noting the method and the outcome. For example, "Attempt: Filtering after pagination. Result: Some matching records were omitted due to only the current page being filtered. Decision: Apply the filter before pagination."

Identify what must not change to maintain the integrity of existing decisions, such as public API behavior, authentication rules, database schema, dependencies, and unrelated UI components. Although these are not guarantees, explicit constraints aid in spotting when a proposed fix exceeds the intended scope. Save this information as AI_HANDOFF.md or in a location where your team records project details.

Include the current task, a checkpoint commit or backup, the specific change, the expected outcome, relevant files, known issues, past attempts, constraints, and what must remain unchanged. Update the handoff as necessary when the work evolves.

Utilize workspace features like knowledge bases, skills, and agents to enhance the project handoff process. A knowledge base can store requirements and decisions, while a skill can define a consistent review procedure, such as checking for regressions and missing tests before suggesting stylistic changes. An agent can guide you through steps like inspecting files, proposing a change, and verifying the results. However, these tools complement rather than replace a recoverable code checkpoint and your review of the diff.

Before accepting any change, review the diff, run the pertinent tests, and test the main workflow. If the changes fail, return to the checkpoint instead of stacking speculative fixes indefinitely. The objective is to ensure each modification is comprehensible, reviewable, and reversible. Create a template for the handoff and complete it for a real task to solidify your understanding of this process.