Configure permissions

Configure Permissions Wizard

Interactive wizard that configures Claude Code permission rules. The goal: reach near-skip-permissions convenience while keeping a safety net for destructive commands.

Overview

This skill asks 2 questions for most users (scope + preset), then writes the permissions config. Only the "Custom" preset goes deeper with additional questions.

Permissions are written to a Claude Code settings JSON file. The skill deep-merges — it only replaces the permissions key and preserves everything else (model, env, skipDangerousModePermissionPrompt, etc.).

Round 1: Scope + Preset (always asked)

Question 1 — Scope

Ask the user where permissions should be saved. Use AskUserQuestion with these options:

Option	File	Description
All my projects	~/.claude/settings.json	These permissions apply everywhere, across every repo
This project, for everyone	.claude/settings.json	Saved in the repo so teammates get the same permissions when they clone it
This project, just me	.claude/settings.local.json	Only affects you on this machine, invisible to teammates

Question 2 — Preset

Ask the user which preset to use. Use AskUserQuestion with these options:

Option	Description
Balanced Dev (Recommended)	Common dev tools allowed, destructive commands prompt before running
Full Trust	Everything allowed, destructive commands still prompt before running
Read Only	Read + web only, everything else prompts
Custom	Configure each category yourself (more questions follow)

If a preset is chosen (not Custom): skip straight to writing the file. No more questions. This is the fast path.

Rounds 2-4: Custom Preset Only

Only proceed with these rounds if the user picked "Custom" in Q2.

Round 2 — Core Permissions (3 questions)

Q3 - File operations (multiselect):

• Read
• Edit
• Write
• NotebookEdit

Q4 - Bash access:

• Common dev tools (allows specific tool patterns — goes to Round 3 for granularity)
• All Bash (allows Bash broadly)
• Ask every time (puts Bash in ask)

Q5 - Web & tasks (multiselect):

• WebSearch
• WebFetch
• Task (subagents)

Round 3 — Dev Tool Granularity (only if "Common dev tools" chosen in Q4)

Q6 - Git operations (multiselect):

• Read-only: git status, git log *, git diff *
• Stage + commit: git add *, git commit *
• Branch management: git branch *, git checkout *, git switch *, git merge *
• Push + pull: git pull *, git fetch *, git push *

Q7 - Build tools (multiselect):

• npm/yarn/pnpm: npm *, npx *, yarn *, pnpm *
• Python: python *, python3 *, pip *, pip3 *, pytest *, uv *
• System utilities: ls *, cat *, head *, tail *, find *, grep *, wc *, mkdir *, cp *, mv *
• Docker: docker *, docker-compose *

Round 4 — Safety + Mode (2-3 questions)

Q8 - Dangerous commands:

• Ask before running (puts dangerous commands in ask — recommended)
• Allow all (no dangerous command protection)

Q9 - MCP tools (only if MCP servers are detected in .mcp.json or ~/.claude.json):

• Allow all (adds all detected MCP tool prefixes to allow)
• Ask each time (adds them to ask)

Q10 - Default for uncovered tools:

• Ask (default Claude behavior — prompts for anything not in allow)
• This is informational only — Claude Code's default is to ask for anything not explicitly allowed

Preset Definitions

Balanced Dev

{
  "permissions": {
    "allow": [
      "Read", "Edit", "Write", "WebFetch", "WebSearch", "Task",
      "Bash(git status)", "Bash(git log *)", "Bash(git diff *)",
      "Bash(git add *)", "Bash(git commit *)", "Bash(git branch *)",
      "Bash(git checkout *)", "Bash(git pull *)", "Bash(git fetch *)",
      "Bash(npm *)", "Bash(npx *)", "Bash(yarn *)", "Bash(pnpm *)",
      "Bash(node *)", "Bash(python *)", "Bash(python3 *)",
      "Bash(pip *)", "Bash(pip3 *)", "Bash(pytest *)", "Bash(uv *)",
      "Bash(ls *)", "Bash(cat *)", "Bash(head *)", "Bash(tail *)",
      "Bash(find *)", "Bash(grep *)", "Bash(wc *)", "Bash(mkdir *)",
      "Bash(cp *)", "Bash(mv *)", "Bash(make *)", "Bash(cargo *)",
      "Bash(bun *)", "Bash(deno *)"
    ],
    "ask": [
      "Bash(rm -rf *)", "Bash(rm -r *)", "Bash(git push --force *)",
      "Bash(git push -f *)", "Bash(sudo *)", "Bash(chmod 777 *)",
      "Bash(chmod -R *)", "Bash(mkfs *)", "Bash(dd *)",
      "Bash(git reset --hard *)", "Bash(git clean -f *)",
      "Bash(> *)", "Bash(curl * | bash*)", "Bash(curl * | sh*)",
      "Bash(wget * | bash*)", "Bash(git push *)"
    ]
  }
}

Full Trust

{
  "permissions": {
    "allow": [
      "Read", "Edit", "Write", "NotebookEdit", "Bash",
      "WebFetch", "WebSearch", "Task"
    ],
    "ask": [
      "Bash(rm -rf *)", "Bash(rm -r *)", "Bash(git push --force *)",
      "Bash(git push -f *)", "Bash(sudo *)", "Bash(mkfs *)", "Bash(dd *)",
      "Bash(git reset --hard *)", "Bash(git clean -f *)"
    ]
  }
}

Read Only

{
  "permissions": {
    "allow": [
      "Read", "WebFetch", "WebSearch", "Task"
    ],
    "ask": [
      "Edit", "Write", "Bash"
    ]
  }
}

MCP Server Detection

Before writing the file, detect MCP servers:

1. Read .mcp.json in the current project root (if it exists)
2. Read ~/.claude.json (if it exists)
3. Extract all server names from the mcpServers key in each file
4. For each server name, the MCP tool permission format is: mcp__<server_name> (double underscore, with hyphens in server names converted to underscores)

For Balanced Dev and Full Trust presets: automatically add all detected MCP server tool prefixes to the allow list. For example, if servers exa and figma-dev are found, add "mcp__exa" and "mcp__figma_dev" to allow.

For Read Only: add MCP server tool prefixes to ask.

For Custom: ask the user (Q9) whether to allow or ask.

Writing the Settings File

Step 1: Read existing file

Read the target settings file (based on scope choice). If it doesn't exist, start with {}.

Step 2: Build permissions object

Based on the preset (or custom selections), construct the permissions object with allow and ask arrays.

Include detected MCP servers as described above.

Step 3: Deep merge

Parse the existing JSON. Replace ONLY the permissions key with the new permissions object. Keep all other top-level keys intact.

Step 4: Write file

Write the merged JSON back to the file, formatted with 2-space indentation.

If the directory doesn't exist (e.g., .claude/ for project-level), create it first.

Step 5: Confirm

Tell the user:

• Which file was written
• How many rules were added to allow vs ask
• Remind them that ask rules still prompt once (they're not blocked)
• Suggest they restart Claude Code for changes to take effect

Building Custom Permissions

When the user picks "Custom", build the allow and ask arrays from their selections across Rounds 2-4:

1. File operations (Q3): Selected tools go to allow, unselected go to ask
2. Bash access (Q4):
   • "All Bash": Bash goes to allow
   • "Common dev tools": specific patterns from Q6+Q7 go to allow
   • "Ask every time": Bash goes to ask
3. Web & tasks (Q5): Selected tools go to allow, unselected go to ask
4. Git operations (Q6): Selected patterns go to allow as Bash(pattern) entries
5. Build tools (Q7): Selected patterns go to allow as Bash(pattern) entries
6. Dangerous commands (Q8):
   • "Ask before running": all dangerous patterns go to ask
   • "Allow all": dangerous patterns are not added to ask
7. MCP tools (Q9): detected servers go to allow or ask based on selection

Important Rules

• Never use deny — deny blocks Claude entirely, even if the user explicitly requests the action. Always use ask instead, which prompts once so the user retains control.
• Deep-merge only — never overwrite the entire settings file. Only replace the permissions key.
• git push in ask for Balanced Dev — regular push (not just force-push) is in ask because pushing is a shared-state action visible to others.
• Presets skip all further questions — if the user picks Balanced Dev, Full Trust, or Read Only, write the file immediately after Q2. Do not ask rounds 2-4.