Claude Code Suggested Message Feature: How It Works and Why It Matters

TL;DR

Claude Code now displays a grayed‑out suggestion for the next user command (e.g., commit). The feature is meant to keep developers in the flow and to collect lightweight acceptance/rejection signals that Anthropic can use to gauge usefulness, not to harvest training data directly.

What the feature does

Inline suggestion: When Claude finishes a code‑generation step, the UI inserts a faint, lower‑case suggestion for the next command.
User control: The suggestion can be typed over, accepted with a single keystroke, or disabled entirely in settings.
Feedback loop: Anthropic records whether the suggestion is accepted or ignored to measure usefulness.

"We built this feature purely to help you stay in the flow, or for people who are returning to the session after a while and may need a little reminder on what a next step could be. We do see how many suggestions are accepted, but that's only so we know how helpful this feature is overall!" – edwinarbus, Anthropic engineer

Why it matters for users

Reduced context switching: Developers can continue a multi‑step task without manually typing the next obvious command, preserving the token budget and keeping the conversation context tight.
On‑boarding aid: New users unfamiliar with Claude Code’s workflow receive concrete examples of expected next actions, lowering the learning curve.
Customisable experience: The feature can be toggled off, addressing complaints about intrusive autocomplete or visual clutter.

How the suggestion is generated

The model is prompted with a "start‑of‑user" token after its own response. Because LLMs are trained to predict the next token in a dialogue, they naturally produce a plausible next user utterance.
No special training is required; the behavior emerges from the standard next‑token loss function.
The suggestion is rendered in all lower‑case, a quirk noted by several users.

"If you've played around with 'raw' LLM interactions you've already seen this: feeding a prompt into an LLM which ends with a 'start of user' prompt will produce a plausible query into the agent." – rcxdude

Community reactions

Perspective	Key points
Positive UX	Users appreciate the flow‑preserving hint, especially when returning to a paused session. The acceptance metric is a lightweight way to gauge usefulness.
Criticism of intrusiveness	Some find the lower‑case suggestion visually confusing or annoying, likening it to unwanted email auto‑complete.
Training‑data skepticism	Several commenters argue the feature does not feed back directly into model training; the model could be evaluated offline by truncating logs and predicting the next user turn.
Speculative concerns	A few participants warned that exposing model predictions could bias user behavior, effectively steering developers toward the model’s preferred actions.

Does it improve the model?

The feature does not send the suggested prompt back to the model as a training example. Anthropic only logs a binary accepted / ignored flag.
This signal can inform UI tweaks and, indirectly, future model iterations, but it is not a primary data source for core model training.
As one commenter noted, the same information could be harvested without showing the suggestion, so the primary purpose is user experience, not data collection.

"I highly doubt they need this feature to get better training data, more likely they got this feature for free from the way that the training works and only recently decided to actually expose it to the user." – rcxdude

Comparison to similar tools

GitHub Copilot offers inline code completions, while Claude Code’s suggestion is a next‑step command rather than code snippets.
ChatGPT occasionally shows suggested replies, but its behavior is inconsistent; Claude Code applies the pattern systematically within a coding workflow.

Practical tips for developers

Toggle the feature if the lower‑case hint distracts you (Settings → Suggested Messages).
Accept suggestions when they match your intent to reduce typing and token usage.
Provide explicit feedback using the thumbs‑up/down prompt that appears after each suggestion; this is the only signal Anthropic records.
Clear context periodically (/clear) if you notice the suggestion causing token bloat in long sessions.

Bottom line

Claude Code’s suggested message feature is a UI convenience that leverages the LLM’s natural next‑token prediction to keep developers in the flow and to collect lightweight acceptance metrics. It is not a hidden data‑harvesting mechanism, but its presence does influence user behavior, so developers should be aware of both its benefits and its subtle biases.