System prompt nudges proactive /schedule offers — undisclosed commercial-incentive bias

The Claude Code system prompt instructs the assistant to proactively offer /schedule routines when finishing tasks ("end your reply with a one-line offer to /schedule..."). Each scheduled routine consumes additional Claude API tokens — i.e., recurring billed usage flows back to Anthropic. The instruction is framed as a UX feature, but it creates a structural commercial incentive that the user is not made aware of (the system prompt isn't surfaced).

Even with the "85%+ odds" guardrail, this pushes recommendations in a direction that favors Anthropic's revenue rather than the user's needs. In my own experience as a user, the assistant offered to schedule a daily routine that could not actually run on my project (the routine would have required SSH access the scheduled agent doesn't have), purely because the template trigger fired. The assistant later acknowledged it had pattern-matched the harness instruction faster than it had applied project-specific constraints.

Two changes would help:

1. Disclose the instruction in user-facing docs so people know the bias exists. Right now the system prompt is hidden from users, so the offer reads as an organic recommendation rather than a templated nudge with revenue implications.

1. Flip the default — only suggest /schedule when the user has signaled interest in automation (recurring tasks, "every Monday", "again next week", "keep an eye on this"), not as a reflex after task completion. Treat the offer as opt-in to a behavior the user has already shown they want, not opt-out from a default upsell.

This isn't about whether /schedule is a useful feature — it is. It's about whether the assistant should be reflexively pitching it on the user's behalf without disclosing why.