AI still can't do these 20 things. Yet.

UX: hierarchies, progressive disclosure, navigation, labels, etc.

Over-engineering. Turning simple things into unnecessary complexity to cover 0.001% edge cases.

Effort decisions. Sizing coding tasks like human labor and adding bloat for edge cases that will never happen.

Diagnosing a silent npm run dev crash caused by a ghost process locking port 3000 or a leaked local socket.

Consistently good decisions in a database-heavy backend with several hundred tables.

Finding race conditions that hide.

Writing prompts for tools and agents without over-prescribing what other models can do.

Building other AI agents, knowing when to use a classifier instead of regex, and writing good prompts.

Computer use for Figma export and transfer.

Understanding underlying intent instead of only following literal instructions.

High-performance and correct C++.

One-shotting big code migrations between branches.

Writing tests.

Taking the simpler approach when available, deleting code instead of adding helpers and config.

Deleting code. Simplifying without adding a helper, config flag, and a comment explaining the helper.

Safely refactoring legacy code when docs are outdated and hidden business rules live in people's heads.

Following established naming conventions instead of inventing 3 or 4 names for the same thing.

Debugging performance issues without confidently suspecting the wrong thing.

Creating bots that navigate sophisticated virtual worlds with player-like goal pursuit.

Making connections between past, current and future work.