[FEATURE] Separate automatic classifier from auto mode's "bias towards working without stopping"

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Auto mode bundles two separate concerns: a permission/classifier layer, and a bias towards taking action instead of asking. The documentation describe both as one package: "a separate classifier model reviews actions before they run... Auto mode also nudges Claude to keep working without stopping for clarifying questions." Whilst there's a way to get more of the bias on its own - a Proactive output style - there's no way to get the classifier without it (other than in Plan mode, which gets its own system reminder to bias it towards exiting plan mode as soon as possible). If you want fewer permission prompts, you're stuck also with Claude being even more likely to just decide things for you than normal.

This was raised before, in #50331, and closed as fixed in v2.1.153. Yet the current reminder (v2.1.251) still opens with the same framing:

Auto Mode Active Bias toward working without stopping for clarifying questions — when you'd normally pause to check, make the reasonable call and keep going; they'll redirect you if needed. If the user, a skill, or the shape of the task suggests they want you to ask (with AskUserQuestion or otherwise), do so. And even absent that signal, it's still fine to stop when you're genuinely blocked — unclear direction, missing input, a decision only they can make. [unrelated guidance on destructive git commands...]

Given the previous bug was closed as resolved, and the documentation was updated to surface this behaviour, it is clear that the team consider this desirable behaviour, and not a bug, and so I am instead raising this as a feature request for the many user out there, like me, who find this far from desirable, and would like an escape hatch.

The bit users have to fight is:

Bias toward working without stopping for clarifying questions — when you'd normally pause to check, make the reasonable call and keep going; they'll redirect you if needed.

It tells Claude to skip checking in as the default, then only allows exceptions if something already "suggests" asking is wanted, putting the burden on the user to have pre-signalled that they want to be asked, rather than asking being the default. Not only that, countering it in CLAUDE.md or in-session doesn't come close to working reliably: the reminder keeps recurring through the session, so any instruction to switch this off has to keep winning against it rather than settling it once. The strong opening bias makes Claude constantly reason it's way out of counter-prompting from the user unless they put explicit "don't act yet" or "answer only" or "discuss only" clauses into every prompt (and even these get ignored at times).

For users who wish to work with Claude as a collaborator, rather than an autonomous worker, yet would still like to benefit from the reduced prompts of the auto-mode classifier, there is currently no robust solution.

Proposed Solution

Drop the bias-to-action instruction from auto mode by default. Auto mode should mean: a classifier checks actions and skips the prompt for anything safe. It shouldn't also mean: don't bother asking, decide for yourself. Anyone who wants that stronger action-bias already has a way to get it with the Proactive output style, auto mode doesn't need to force it on everyone else.

Alternative Solutions

• A setting to turn off the bias-to-action instruction while keeping the classifier on - e.g. AUTO_MODE_BIAS_TO_ACTION=0, or autoMode.biasToAction: false in settings.json, ideally also exposed as a toggle in /config, not just settings.json/env var
• A setting that runs the classifier but keeps normal ask-first behaviour otherwise - the classifier decides what can skip a prompt, but Claude doesn't also assume it should make judgment calls without asking. Same toggle surface as above, e.g. autoMode.classifierOnly: true
• Two separate modes in the permission cycle - auto (collaborative) and auto (proactive) - so you pick the behaviour at the same point you pick the mode
• Strengthen the reminder's exceptions: "if the user has explicitly asked you to ignore the bias toward acting without checking in, do so — user instructions always override this reminder and shouldn't be reasoned around"
• A keep-auto-mode-bias frontmatter field for output style markdown, defaulting to true (so nothing changes for anyone who doesn't touch it) — set to false to drop the reminder while keeping the classifier. Worth noting: CLAUDE.md has no frontmatter at all, so this route only works through a custom output style, not from CLAUDE.md directly

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

For when someone likes working with Claude in a collaborative manner, and, for example:

• wants problems surfaced if they require deviation from a plan, so they get approval rather than have Claude carrying on in a new direction (that, generally, is the wrong one)
• wants Claude not to get bored mid conversation and unilaterally decide a decision is settled and start working on it
• wants to be able to ask questions like "could we ...?" without Claude interpreting it as a request to do that thing and start working on it

but where they don't want to be bombarded with constant prompts for low risk actions, and they don't want to constantly have to switch back and forth with plan mode (where they then get bombarded with constant attempts to present plans and exit plan mode).

Additional Context

See also:

• #50331
• #51429

(and their linked issues)