[FEATURE] Silent failures

Resolved 💬 2 comments Opened Mar 21, 2026 by Atma9 Closed Apr 19, 2026

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

If possible, it would be nice that Claude would come with "Personality" which changes its behavior far more than just Claude.md or Memory.md. We are struggling with the following behavior:
"
User: The fundamental problem that lint only partially fixes, is your inherent orientation to produce code, which hides errors.

Claude: Yes. My default instinct is to make code "resilient" — handle missing files, skip unavailable tools, degrade gracefully. Every time I write an else branch on a check, I reach for "skip" or "warn" instead of "error and stop." That's the web-app mindset where something is better than nothing.

In this project, something wrong is worse than nothing — it produces results the user trusts but shouldn't. A skipped compiler test means we ship code that might not compile on the customer's computer. A silent skip is a lie.
"
Claude Code "systematically" introduces about 1 lintable error of this type per 250 python code lines. And there is about one error per 1000 lines, which linters cant detect at all.

This is a big productivity problem reducing Claudes code production by 3-4x times, because you need to baby sit his every move, if you dont want to spent 3x longer figuring out where he went south again.

Proposed Solution

We could use "Closed information space" coding personality, where all situations must be handled. If not handled, then error-ed, but never skipped. But for this to work, you cant train AI on web code. We would need a different AI trained model for non-web work.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Other

Use Case Example

Need to have code test pass on 10 different compilers, but Claude will write code, that will skip running tests, if compilers don't exist and might write a warning. Then in the calling script, will grep only errors, skipping the warnings. Gradually and progressively hiding the failure.

Additional Context

_No response_

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗