SuperThink

Deep-dive research and analysis engine. Triggered by /superthink. Runs an interrogate flow to narrow topic and scope, then executes a 4-stage fully automatic...

This skill appears to do what it says: it needs an Anthropic API key and writes large generated documents to persistent storage. Before installing: (1) only provide ANTHROPIC_API_KEY if you intend to run large batch jobs; (2) be cautious about setting NOTIFY_WEBHOOK_URL or TELEGRAM_* vars — those will send pipeline outputs to external endpoints you configure; (3) ensure /data/superthink is an appropriate, isolated place to store potentially sensitive outputs and that you have disk quota for ~40–60k word files; (4) note the pipeline runs fully unattended after you confirm scope — if you want intermediate approvals, modify the cron/poller as described; (5) review or sandbox the required batch-worker and md2docx implementations if you plan to run them in a shared environment. If you want, I can list the exact files the skill will create or suggest minimal sandboxing/permissions to limit exposure.

Type: OpenClaw Skill Name: superthink Version: 1.0.1 The 'superthink' skill implements a highly autonomous, multi-stage research pipeline that operates unattended for 6–12 hours using system cron jobs for state persistence and stage transitions. While the logic in SKILL.md and the supporting scripts (batch-worker.py and md2docx.py) appears aligned with the stated purpose, the skill requires high-risk permissions, including sensitive environment variables (ANTHROPIC_API_KEY, TELEGRAM_BOT_TOKEN), broad filesystem access, and the ability to send data to user-configurable network endpoints (NOTIFY_WEBHOOK_URL). The 'Cron Poller Pattern' is particularly sophisticated and risky, as it involves the agent dynamically creating, updating, and deleting its own scheduled tasks, which could be repurposed for persistence or unauthorized execution if the underlying prompts are compromised.