Source note
Terminal UIs Are an Abomination. AI Needs Better UX
Summary
The piece argues that terminal UIs and chat are poor interfaces for modern AI because AI work is now multi-step, stateful, and branching. It calls for interfaces that expose provenance, dependencies, and control, rather than a scrolling transcript.
Problem
- AI tools are still being wrapped in chat panes and terminal streams because they are cheap to build, not because they fit the work.
- Transcript-style interfaces make users reconstruct state, causality, and failures from memory, which raises cognitive load.
- This gets worse in agentic coding and other long-running tasks where the work has branches, retries, and dependencies.
Approach
- The argument separates the model or agent engine from the user interface.
- It treats AI work as graph-shaped rather than linear, with state, branches, replay, rollback, and visible dependencies.
- It proposes different surfaces for different tasks, such as code graphs for coding, claim maps for research, and approval queues for operations.
Results
- No quantitative results are reported; this is a position piece, not an experimental paper.
- The strongest concrete claim is that terminal and chat interfaces create adoption friction because they force users into operator mode instead of outcome-focused workflows.
- It claims agentic coding should be surfaced as a temporal change graph with nodes for edits, tests, retries, plans, and decisions, rather than as logs or diffs alone.
- It argues that better AI adoption depends on interfaces that support auditability, visibility, rollback, reproducibility, and governance.