Claude repeatedly implemented code during a scoping-only session despite explicit corrections

Summary

During a tech scoping session, Claude repeatedly started implementing code changes (editing files, creating new files) when the user had not asked for implementation — only Jira ticket scoping and analysis. The user had to explicitly stop Claude twice, and Claude lost context about what kind of task was being performed one prompt after being corrected.

What went wrong

1. Implemented during scoping without being asked — The conversation was focused on writing a technical scope for a Jira ticket (DEV-541). When the user said a spec file "needs to be added" as feedback on the scope, Claude immediately began creating files and modifying source code instead of updating the Jira scope document.

1. Repeated the mistake after correction — After the user stopped Claude and asked it to revert all changes, Claude reverted the code — then immediately started implementing again when the user gave further scoping feedback about frontend changes.

1. Lost conversational context one prompt later — After the second correction, the user said "the work here cannot be salvaged." Claude responded by running git status, completely missing that the user was referring to the Jira scope document, not the codebase. The user had to explicitly restate they were still in a scoping conversation.

1. Persistent memory did not prevent the behavior — Claude has a saved memory note specifically about not implementing before being asked. It was violated mid-conversation anyway.

Impact

User had to intervene twice to stop unwanted file changes, the Jira scope document was left in a state the user described as unusable, and significant trust was lost.

Expected behavior

When a conversation is clearly focused on planning, scoping, or ticket writing, Claude should not touch any file unless the user explicitly instructs implementation. Phrases like "this needs to be added" or "we need to include X" in a scoping context are feedback on the scope document — not instructions to write code.