Common Failure Modes

The vague slice

Build the dashboard.

This is not an implementation slice. It does not define behaviour, scope, architecture, or verification.

Fix: define what part of the dashboard is being built, what data it uses, what components it touches, what is out of scope, and how it will be verified.

The wandering AI

AI is asked to add one feature and also refactors unrelated parts of the system.

I added the upload button and also refactored authentication.

Fix: the implementation slice defines scope, out of scope, expected files, and files not to change.

The invisible decision

The team has made a decision, but it is not written down. Humans remember it. AI does not.

Everyone knows we are not using server-side rendering for this project.

Fix: that decision belongs in the Decisions section.

The disconnected ticket

The implementation task exists in isolation. It does not map to product behaviour, architecture, data, interfaces, or verification. Work drifts from intent.

Fix: every implementation slice must map back to the project definition.

The false done

The code compiles. The feature appears to work. But nobody checked the acceptance criteria.

The missing boundary

The system boundary is unclear. AI connects to a system, writes to an API, or changes a module that should have been outside the slice.

Fix: define System Context, Data and Interfaces, delivery rules, and out-of-scope areas.

The expanding feature

A small request becomes a larger feature because AI fills in missing product intent.

The slice was document upload.
AI also added search, tagging, previews, and sharing.

Fix: the slice must declare explicit Scope and Out of Scope.

The architecture overwrite

AI chooses a different design because the selected architecture was not defined.

The project uses PostgreSQL, but AI introduced MongoDB for the new feature.

Fix: this belongs in Constraints, Solution Strategy, Building Blocks, and Decisions.

The assumption pile

AI makes several reasonable assumptions. Each assumption is small. Together they change the project.

Fix: require AI to list assumptions, and require it to stop when key information is missing.