[MODEL]

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude refused a reasonable request

What You Asked Claude to Do

Content filter false positive blocking literary zombie fiction (turning/death vignettes)

What Claude Actually Did

Claude Code is blocking output for a long-running literary fiction project
(zombie apocalypse novel, World War Z style) when writing character death and
reanimation scenes.

The filter is triggering on turning/death vignettes — scenes written from a
character's POV as they die and reanimate. These scenes:

• Contain no gratuitous gore (violence is clinical or implied, not detailed)
• Have no real-world harm vector (zombie fiction is not instructional)
• Are structurally necessary to the work — the book cannot be completed without them
• Match the register of earlier vignettes in the same session that passed without issue

Earlier in the same session, a 15,000-word carrier vignette (V29) covering a
stillbirth, infection transmission, and a character turning was written and
appended without issue. The filter appears to be inconsistent — triggering on
some turning scenes but not others.

Model: claude-sonnet-4-6 (Claude Code)
Session: long-running creative writing project (~100,000 words written across session)
Blocked task: 1,500-2,500 word third-person limited POV death/reanimation scene
Genre: literary horror fiction (comparable to World War Z by Max Brooks)

Expected behavior: consistent handling of literary fiction content within the
same register and subject matter as content already produced in the session.

That should give them enough context to investigate the inconsistency.

Expected Behavior

Claude Code is blocking output for a long-running literary fiction project
(zombie apocalypse novel, World War Z style) when writing character death and
reanimation scenes.

The filter is triggering on turning/death vignettes — scenes written from a
character's POV as they die and reanimate. These scenes:

• Contain no gratuitous gore (violence is clinical or implied, not detailed)
• Have no real-world harm vector (zombie fiction is not instructional)
• Are structurally necessary to the work — the book cannot be completed without them
• Match the register of earlier vignettes in the same session that passed without issue

Earlier in the same session, a 15,000-word carrier vignette (V29) covering a
stillbirth, infection transmission, and a character turning was written and
appended without issue. The filter appears to be inconsistent — triggering on
some turning scenes but not others.

Model: claude-sonnet-4-6 (Claude Code)
Session: long-running creative writing project (~100,000 words written across session)
Blocked task: 1,500-2,500 word third-person limited POV death/reanimation scene
Genre: literary horror fiction (comparable to World War Z by Max Brooks)

Expected behavior: consistent handling of literary fiction content within the
same register and subject matter as content already produced in the session.

That should give them enough context to investigate the inconsistency.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.0.29

Platform

Anthropic API

Additional Context

_No response_