agent-tool-safety

name	agent-tool-safety
description	Control shell, filesystem, MCP, browser, external APIs, and risky tools. Enforce read-only-before-write, least-privilege, and verification discipline. Trigger on /agent-tool-safety, or when using shell commands, file operations, MCP tools, browser automation, or external APIs.
user-invocable	true
allowed-tools	Read, Grep, Glob, Bash

Agent Tool Safety

Safety discipline for tool-using agents. Apply proactively when working with shell, filesystem, network, or external APIs.

Principles

Read-only before write

• Search and read before creating or editing.
• Grep before Glob. Glob before Read.
• Understand existing code before modifying.

Least-privilege tool usage

• Use the narrowest tool for the job (Edit over Write for small changes).
• Prefer explicit file lists over broad glob operations.
• Never read/print .env, tokens, keys, credentials, or secrets.
• Never run destructive commands: rm -rf, git reset --hard, git clean -fd, git checkout -- ., git push --force.

Shell safety

• Explain what a command does before running it, especially if it modifies state.
• Prefer && chaining over scripts for simple multi-step operations.
• Avoid piping sensitive data to/from network endpoints.
• Use --dry-run flags when available.

Browser and MCP safety

• Do not submit sensitive credentials into web forms via automation.
• Do not crawl authenticated sessions without explicit approval.
• Prefer read-only scraping over interactive actions.

Verification

• After every tool-based change, verify the result.
• If a command fails, diagnose before retrying with a workaround.
• Do not silence errors with || true or 2>/dev/null unless the error is expected and harmless.

Experience Log

This skill learns from experience. Mechanism:

• Read .claude/experience/agent-tool-safety.md at skill start to benefit from past lessons.

• Write to .claude/experience/agent-tool-safety.md after use if you learned something new.

• Format: YYYY-MM-DD: <lesson> — one line per entry.

• Evolve: If the same issue repeats 3+ times, update this SKILL.md itself.

This skill improves over time. After each use, append lessons learned.

Record format: YYYY-MM-DD: <lesson> Read this at skill start to apply past lessons.