Claude Code’s suggested message feature: I think the real customer is the model
Anthropic’s Claude Code now suggests the user’s “next message” directly in the input box, prompting debate over whether the feature primarily serves user productivity or model improvement. Commenters question if showing autocomplete-like prompts biases human input, degrades future training data, or simply nudges people to burn more tokens, while others find it genuinely helpful for staying in a coding flow. An Anthropic engineer states the feature is meant only as a UX aid and not for collecting preference signals, but skepticism remains around data use, interface intrusiveness, and the broader trend of AI-driven “helpfulness” encroaching on human communication.
Perceived Purpose of Prompt Suggestions
- Many readers accept the article’s idea that the real “customer” is the model and this is about learning next steps.
- Others argue the same training could be done server‑side by truncating conversations and predicting the next user message, without showing suggestions at all.
- Some suggest the main goal may be user engagement or onboarding: nudging novices to “do something” next, or helping people resume a session.
- A self‑identified Claude Code team member says suggestions are not used for preference training, only for UX/flow and to measure how often they’re accepted.
User Experience and Distraction
- Several people like the feature in Claude Code, especially for quickly confirming “yes, do that / commit” during implementation phases, where tab/enter is faster than typing.
- Others immediately turn it off, calling it distracting, misleading, or “self‑fulfilling”: they sometimes accept a suggestion they didn’t really want.
- Users with ADHD or similar issues say merely having to ignore suggestions is a cognitive tax and breaks focus.
- Some note it’s subtle and non‑blocking; others complain about visual confusion (e.g., greyed text indistinguishable in some terminals).
Training Data, Feedback, and Ethics
- One technical line of argument: showing a suggestion and seeing whether it’s accepted provides a clean RL signal (accept = good, reject = bad).
- Counter‑argument: suggestions bias user inputs, making them less independent and potentially degrading training data quality (“AI slop prompting itself”).
- There is concern about labs promising not to train on paid prompts; some fear features like this are backdoor normalization of training on user data.
- A few users openly talk about intentionally poisoning thumbs‑up/down feedback because they resent doing unpaid labeling work.
Style, Communication, and Cultural Norms
- Some dislike that suggestions use all‑lowercase or SMS‑style abbreviations they’d never use, finding it socially jarring or rude.
- Broader backlash against AI‑written replies in email and messaging: many feel these cheapen human relationships and reflect product teams chasing “AI usage” metrics.
Business Models and Product Strategy Concerns
- Some speculate about future per‑token pricing and incentives to burn more tokens (longer contexts, more steps).
- Others see this and similar “copilot everywhere” features as examples of big companies flailing, shipping low‑value AI add‑ons to satisfy internal mandates rather than user needs.