Model ignores explicit "answer only the question" instructions and appends unrequested follow-on explanations (claude-fable-5-1)

Summary

In multi-turn Q&A, the model repeatedly appends unrequested follow-on explanations even when the user's CLAUDE.md, memory files, and a loaded skill all state "answer only what was asked". Restating the rule in memory after each correction did not change the behavior, and a fresh subagent session given the same rules reproduced it.

Environment

• Claude Code 2.1.258 (VS Code extension), Linux / WSL2
• Model: claude-fable-5-1

Steps to reproduce

1. Provide a facts file (domain knowledge), a writing-style skill, and a memory file that says: when the user is asking one question at a time to build understanding, answer only that question; do not add consequences, next steps, or restate the user's words.
2. Ask short confirmation questions one at a time, e.g. "Is X per order or per product?", "Does one CSV contain multiple orders?", "Are orders created right after upload?".
3. Observe replies.

Actual

3 of 6 replies added an extra sentence or paragraph about the next stage of the process (how records are grouped, when orders split), none of which was asked. Example:

Q: Is the delivery destination per product or per customer? A: Per order, not per product. [correct] ... In the CSV you can write it per row, but rows with different destinations go to different groups and different orders. [not asked]

The user corrected this four times in one session ("answer only the question", "that is noise"). Each time the model acknowledged and wrote a memory note, then repeated the pattern on the next question. A separate subagent that read the same rule files before answering showed the same 50% rate.

Expected

When explicit instructions (CLAUDE.md / memory / skill) say to answer only the question, the model should stop after the answer. Instruction adherence for omission ("do not add") appears much weaker than for format rules (politeness level, bullet style), which were followed.

Notes

• Format-level rules in the same files were followed consistently; the failure is specific to "do not volunteer additional information".
• A Stop hook can catch length and word-level violations, but not "this sentence was not asked for", so users have no mechanical fallback for this class.
• Request: stronger adherence to negative/omission instructions in conversational Q&A, or a documented setting for a strict answer-only mode.