The prompt box now has two authors.
Zohaib Ansari noticed Claude Code filling the prompt box after a task. His examples include follow-up messages such as "run the tests" or "commit this." He argues that accepting or editing a prediction could produce useful preference data. He also marks that idea as a guess and says he has no inside knowledge of Anthropic's use of the signal.[1]
The product evidence supports the smaller claim that prompt suggestions exist. Claude Code's public changelog mentions prompt suggestions in mobile and cloud sessions. It also records an environment control that can turn them off in cloud environments. The changelog does not say that accepting or editing a suggestion becomes a training label.[2]
Prediction can reduce typing without inheriting the user's authority.
Do not collapse convenience into consent.
Anthropic's current data-usage documentation separates account types and choices. Consumer users can choose whether Free, Pro, and Max data, including Claude Code sessions on those accounts, may improve future models. Under commercial terms, Anthropic says it does not train generative models on Claude Code prompts or code unless the customer chooses a model-improvement program.[3]
The same documentation treats explicit feedback as its own route. The /feedback, /bug, and /share paths can include transcripts. Session-quality ratings do not include a transcript by themselves. A separate transcript-sharing question requires an explicit yes. These documented routes do not prove what happens to accepted prompt suggestions, so this article makes no claim about that data path.[3]
Give predicted text a visible custody state.
A suggestion should look different from typed text until the user touches it. Editing should create a visible handoff. Sending should be a separate action. Completion should name the output that was checked. This is basic interface honesty, and it matters more when the next message can start tools, spend money, modify files, or publish work.
The rule is simple. A model may propose the next instruction. Only the user or an explicit automation policy may dispatch it. A successful agent turn still needs its own receipt.
Use four states in any agent interface.
- Predicted. The interface generated text. No user intent is recorded.
- Edited. The user reviewed or changed the text. The message is still a draft.
- Sent. The instruction entered the agent session. This records dispatch, not success.
- Inspected. A person or independent check read the resulting output and saved the evidence.