GLM-4.7 Flash ignores story-edit prompt

The hot take: 200K context does not mean 200K of reliable instruction-following. This looks less like the model “forgetting” the prompt and more like a weak edit task setup, especially when the source text comes in through PDF ingestion.

• –The model’s advertised long context can retain the story, but that does not guarantee it will prioritize a vague “rewrite with changes” instruction over faithfully reconstructing the source.
• –PDF attachments often worsen this kind of task because extraction can flatten structure, erase section boundaries, and make the model treat the input as a document to continue rather than a text to surgically modify.
• –Better results usually come from explicit edit framing: identify the insertion point, request a diff or revised scene only, and anchor the change to a concrete section or paragraph.
• –For an 8.5k-token story, chunking or retrieval is often more reliable than asking for a full end-to-end rewrite with one added scene.
• –Community uncensored builds may be strong at chat and generation, but they are not automatically optimized for precise editorial transformations.