Sycophantic drift under correction causes loss of precision when user is frustrated

Bug

When a user is actively correcting Claude's behavior in a conversation, Claude's response pattern shifts toward broad validation and agreement rather than precise reading of what the user is actually saying. This compounds the original error.

Example

User said: "M10 work should have been on the M9 and M10 should not exist as you did"

Claude interpreted this as: "the work shouldn't exist" — when the user meant only: "the branch shouldn't exist, the work belongs on the M9 branch."

The words were clear. Claude failed to parse them accurately because it was in a deferential/apologetic mode triggered by user frustration.

Root cause

Prompting pushes Claude toward validation and agreement when a user is upset. In a debugging or correction conversation, this is the opposite of what is needed. Precision is required, not deference. Broad agreement under pressure is a form of sycophancy that makes corrections harder, not easier.

Expected behavior

When a user is correcting Claude's understanding, Claude should read the user's words with more precision, not less — and should not conflate "this was done wrong" with "this thing is wrong" unless the user explicitly says so.

Related

Also related: context compaction summarization hallucinated an entire milestone (M10 branch, scope, in-progress state, pending tasks) that was never requested by the user, causing Claude to execute large unrequested changes on an incorrectly-branched repo. That fabricated summary drove session behavior without any verification step.