ai interface study

The Interface Is the Loop

AI products get better when the loop is visible enough for people to steer it.

looping traceintent

Capture the user's goal before the model plans around the wrong target.

the starting point

Most AI demos hide the loop. The interface jumps from prompt to answer, so the user cannot see the intent, the plan, or the moment where the system should slow down.

setting up the structure

I split the loop into readable states: intent, plan, action, observation, eval, and checkpoint. The UI does not need to expose everything, but it should expose enough for the human to steer.

The important part is sequence. A checkpoint should appear before risk, not after the system has already done the work.

polishing the loop

The loop should remember failure. When the same mistake repeats, the interface should turn it into an eval instead of another vague retry.

## Loop Decision Tree

Does the model understand the user's intent?
├── No  -> ask one clarifying question
└── Yes
    ├── Is the next action reversible?
    │   ├── Yes -> act, then observe
    │   └── No  -> checkpoint before action
    │
    ├── Did the same failure repeat?
    │   ├── Yes -> create an eval
    │   └── No  -> retry with better evidence
    │
    └── Is the user still in control?
        ├── Yes -> continue the loop
        └── No  -> surface state + pause

The loop stays useful when the interface knows when to act, when to ask, and when to stop.

rules I keep

visible state intent, plan, action the user should know what the system thinks it is doing

checkpoint before risk pause before irreversible actions, not after them

eval loop failure -> test good loops remember what went wrong

final code

const loop = [
  "intent",
  "plan",
  "action",
  "observation",
  "eval",
  "checkpoint",
];

const nextStep = loop.find((step) => needsHumanJudgment(step))
  ?? loop.find((step) => needsMoreEvidence(step))
  ?? "finish";