review-permissions

Invoke this Skill at the end of a session to review commands that required manual approval. It reads ~/.claude/settings.json, identifies gaps, and outputs a proposed diff with broad wildcard patterns like Bash(git:*) instead of individual subcommand entries.

Use acceptEdits mode to auto-approve file operations, then maintain a consolidated Bash allow list with one wildcard per tool family. This Skill analyzes your session and recommends which broad patterns to add so repeated commands stop triggering prompts.

Yes, it checks permissions.deny before recommending additions and never suggests entries that conflict with the deny list. When adding a broad pattern like Bash(git:*), it also flags destructive subcommands such as git reset --hard that need deny list entries.

There is no formal approval log, so recommendations rely on commands executed in the current session and the agent's memory of permission prompts. Invoke the Skill while the conversation context is fresh for accurate results.

The Skill targets Claude Code specifically, as declared by its supported-harnesses metadata. It operates on the ~/.claude/settings.json permissions structure, which is specific to the Claude Code configuration format.