Prompt suggestion system prompt contradicts CLAUDE.md customization

Summary

The prompt suggestion (autosuggest/ghost text) system prompt sent to the model says:

"Your job is to predict what THEY would type - not what you think they should do."

Users who customize their CLAUDE.md to instruct the suggestion fork to act as an advisor (suggesting what the user should type, not just what they would type) cannot override this system prompt. The CLAUDE.md instruction is seen by the fork model but is contradicted by the injected system prompt.

The conflict

• User's CLAUDE.md: "Suggest what Blake should type next — the most helpful next step — not just what he might type based on patterns."
• Claude Code system prompt: "Your job is to predict what THEY would type - not what you think they should do."

These are literally opposite instructions. The system prompt wins because it's injected by the client at a higher priority than user-provided CLAUDE.md context.

Suggested fix

Either:

1. Allow users to customize the suggestion system prompt (e.g. via settings.json or a promptSuggestionMode: "advise" | "predict" setting)
2. Change the default prompt to be advisory rather than purely predictive — "suggest what would be most helpful for the user to type next" is more useful than "predict what they would type"
3. At minimum, don't actively contradict CLAUDE.md instructions — if the user's CLAUDE.md says "advise not predict", the system prompt shouldn't say "predict not advise"

Context

Discovered during a multi-agent Claude Code session where the user noticed the autosuggest was pattern-matching their typing habits rather than acting as a copilot/advisor. The user explicitly authorized sharing this finding.

Environment

• Claude Code CLI
• Multiple concurrent sessions with shared CLAUDE.md