review-permissions

name: review-permissions description: Audit the current conversation for every permission prompt Claude generated, explain how to suppress each one in settings.json, and recommend whether to do so.

Review Permissions Skill

Scan the current conversation for every instance where Claude asked the user to approve a tool call, then produce a structured audit with suppression instructions and a recommendation for each.

1. Identify Permission Prompts

Go through the full conversation from the beginning. For each tool call that required user approval, record:

• Tool name — the exact tool (e.g. Bash, Write, mcp__github__push_files)
• Command / arguments — the specific invocation that triggered the prompt (e.g. brew install foo, git push --force)
• Why it was prompted — was it missing from the allow list, or does it match a pattern in deny?

Group duplicates (same tool + same command pattern) as a single entry. Do not list tool calls that were auto-approved.

2. Explain the Suppression Config

For each identified permission prompt, show the exact JSON snippet to add to the permissions.allow array in ~/.claude/settings.json (source: ~/dotfiles/claude/settings.json) that would suppress it in future.

Use the correct pattern syntax:

• Bash commands: "Bash(command subcommand:*)" — include subcommands if needed for precision, use * wildcard for arguments
• Specific tools: "ToolName" (e.g. "Write", "Agent")
• MCP tools: "mcp__server__tool_name" or "mcp__server__*" for all tools on a server
• Named tool variants: use the most specific pattern that covers the use case without over-permitting

Show both the minimal (precise) pattern and the broader wildcard pattern where they differ, so the user can choose the scope.

3. Recommend: Add or Keep Prompt

For each entry, make a clear recommendation: Add to allow or Keep as prompt, with a one-sentence rationale.

Use this heuristic:

• Add to allow if: the operation is low-risk, reversible, idempotent, read-only, or routine for this type of project
• Keep as prompt if: the operation is destructive, irreversible, affects shared state (remote pushes, external services), involves credentials, or should always require conscious intent

Be direct — err toward fewer prompts for clearly safe operations and toward keeping prompts for anything with real blast radius.

4. Present the Audit Table

Output a markdown table with these columns:

Follow the table with a summary count: X of Y prompts recommended for suppression.

5. Await User Decision

After the table, ask the user which entries (by number, range, or "all recommended") they want to add to the allow list.

Do not modify settings.json until the user responds.

6. Apply Selected Changes

For each entry the user approves:

1. Read ~/dotfiles/claude/settings.json.
2. Add the config snippet to the permissions.allow array (maintaining alphabetical or logical grouping where possible).
3. Do not add duplicates — check if an equivalent pattern already exists.

After all edits, run:

~/dotfiles/install.sh $(cat ~/.dotfiles_profile)

Confirm which entries were added and which (if any) were skipped as duplicates.