Feature: User-contributed training priors

Feature Request: User-Contributed Training Priors

The Problem

Claude's memory system is scoped per-user, per-project. When a user discovers a hard-won lesson through real-world experience — like "App Store review always flags BYOK API key apps under Guideline 2.1(b), so proactively include a demo key and business model explanation in review notes" — that learning only benefits that one user going forward. Every other user with a BYOK app will hit the same rejection, waste the same 1-2 weeks on back-and-forth with Apple, and teach Claude the same lesson independently.

The Scenario That Triggered This

I built an iOS app that uses a bring-your-own-key model for AI features (users paste their own API keys from Google/Anthropic/OpenAI). Claude helped me prepare and submit the app to the App Store. Claude knew that Apple flags API key fields under Guideline 2.1(b) — but didn't apply that knowledge proactively during submission. The app was rejected, Apple asked for a business model explanation and a demo API key, and the review cycle cost 11 days. The fix was a two-paragraph reply.

When I pointed this out, Claude acknowledged it should have caught it. I asked Claude to save it to memory so it wouldn't happen again — and it did, but only for me. The next developer who asks Claude to help submit a BYOK app will hit the exact same wall.

The Proposal

Add a mechanism for users to flag specific learnings as candidates for model-wide improvement. Something like:

• A "promote to training" action in Claude Code or the conversation UI
• User describes the lesson and why it's broadly applicable
• It enters a review/curation pipeline at Anthropic
• If validated, it becomes part of Claude's baseline behavior for all users

This is different from existing feedback channels:

• Thumbs up/down rates response quality but doesn't capture domain knowledge
• GitHub issues report bugs in Claude Code the tool, not gaps in Claude's applied knowledge
• Memory fixes it for one user but the lesson dies there

Why This Matters

The highest-value knowledge isn't on the internet — it's in the gap between "Claude knows this fact" and "Claude applies this fact at the right moment." Users are already discovering these gaps and teaching Claude the fixes. There's just no path from an individual lesson to a universal improvement.

Users shipping real software are the best source of applied, contextual knowledge. A structured contribution path would turn every user's hard-won lesson into baseline capability for everyone.