Use when user asks to leverage claude or claude code to do something (e.g. implement a feature design or review codes, etc). Provides non-interactive automat...
Security Analysis
medium confidenceThe skill's purpose (driving the Claude Code CLI) matches its instructions, but it explicitly recommends bypassing permission prompts and auto-accepting edits which materially increases the risk of unreviewed file modifications when the agent runs — proceed only after reviewing these behaviors and running in an isolated environment.
Name/description, examples, and runtime instructions consistently target the Claude Code CLI and developer workflows (read/write/edit, run tests, use tmux). No unrelated credentials, binaries, or config paths are requested.
SKILL.md instructs the agent to run arbitrary claude CLI commands in project workdirs, read and write project files, tail logs, compute hashes, and persist registry state. Critically it recommends using --dangerously-skip-permissions and --permission-mode acceptEdits to avoid permission prompts and auto-accept edits — this directly reduces human oversight and can cause unreviewed code changes. The instructions otherwise stay within a coding-agent scope but explicitly enable unsafe, non-interactive behavior.
No install spec in the skill bundle (instruction-only), but SKILL.md tells users to npm install -g @anthropic-ai/claude-code (global npm install) and to have tmux. Suggesting a global npm install requires elevated privileges and should be validated (official package, correct source). No download-from-arbitrary-URL instructions in the bundle.
The skill declares no required env vars or credentials, which matches its stated purpose. Examples reference external tools (e.g., mcp__datadog, --mcp-config) that could imply needing monitoring credentials at runtime — those are not declared here and would be provided by the host environment. This absence is acceptable but users should be aware that some example usages require separate credentials not managed by the skill.
always:false (good) and autonomous invocation is allowed (platform default). The skill instructs storing persistent polling state in a registry (update_registry) and recommends bypassing permission prompts and auto-accepting edits — when combined with autonomous invocation, these behaviors increase the blast radius by enabling unattended, persistent code modifications. The skill does not request system-wide config changes, but its recommended flags reduce safeguards.
Guidance
This skill is coherent for automating the Claude Code CLI, but it explicitly recommends running in non-interactive modes that bypass permission prompts and auto-accept file edits. Before installing or enabling it: 1) Confirm you trust the claude CLI package and install it in a controlled way (prefer non-global install or containerized environment). 2) Avoid using --dangerously-skip-permissions and --permission-mode acceptEdits unless you run the agent in an isolated repo/VM and accept unattended changes. 3) Restrict allowedTools to the minimum necessary (do not grant Write/Edit broadly). 4) If you plan autonomous runs, require explicit human approval for high-impact tasks and audit the registry/logs the skill will write. 5) If examples reference monitoring or other services (mcp__datadog), ensure those credentials are provisioned separately and scoped narrowly. If you want, I can list exact flags and file paths in this SKILL.md that you should disallow or require confirmation for.
Latest Release
v0.1.0
- Initial release of claude-skill. - Provides a managed automation layer for running Claude Code as a coding agent, including worktree setup and PR merging. - Supports non-interactive ("quick mode") and full interactive workflows, with robust logging and adaptive polling for long-running tasks. - Offers CLI flags to control permissions, select models, restrict tools, and customize output. - Includes resilience against CI/container PTY issues, output logging strategies, and polling-based timeout management to prevent premature termination of tasks.
More by @feiskyer
Published by @feiskyer on ClawHub
