AI Compass / Knowledge cards

A second opinion before risky commands

yes/no guards

Checked 11 October 2026 · Michael Engel · Automation · Prompting · Costs

A hook is a script that Claude Code runs automatically at fixed points, for example before a command runs or before the agent stops. If a decision model is attached to that hook, it answers a yes/no question, and Claude has to abide by the answer. This way a different model from the one doing the work does the checking.

What it is for

An agent that works without asking back gets a second opinion on risky commands, rule violations and a premature “done”, without a person having to sit next to it.

Steps

1. Choose the question: May the command run? Does the change violate a project rule? Is the task demonstrably done?
2. Check whether a fixed rule without a model is enough, for example a list of forbidden commands in the hook. If so, use that.
3. Clarify what the checker gets to see. Commands, conversation history or file contents go to a model provider.
4. Read third-party hook code before installing it, and at first use it only in a test project.
5. Set a confidence threshold, that is, the minimum certainty at which the answer counts. In the video, Belay uses 75%. Turn on the decision log.
6. Watch for false alarms. In the test, Toolgate stopped a dangerous delete command, but also deletions that were intended.
7. After the test phase, decide whether the guard stays, is adjusted, or goes again.

Tools from the video

• Toolgate: stops risky commands. In the test it rated a delete command at “77% destruction risk” and blocked it, even though Claude was running without asking back.
• Abide: builds a rule list from CLAUDE.md and AGENTS.md and checks every change against it. Claude has to fix every objection before it stops.
• Belay: reads the history, looks for evidence of “done”, and sends Claude back if there is none.
• Five more tools (browser tests, message steering, hints for compacting the context, file pre-selection, SEO report) are only briefly introduced in the video.

Limits

• All eight tools come from third parties, not from Anthropic or TypeSafe. All that has been verified is that Toolgate exists as a repository and how it describes itself.
• The figures (77%, 75%, tokens saved) come from individual test runs by the channel.
• The video names no costs. Each call costs money at the model provider, and the amount depends on the provider.

Open alternative: fixed hook rules without a model, or Claude Code’s built-in auto mode. Neither sends content to a further provider.

Links

• Decision models: what a model that only decides can do, and what it costs.
• Checkpoints and gates: Belay is an automatic done-check, so a small gate.
• Maintaining your Claude Code setup: Every hook is part of the setup and needs the same upkeep.
• “Done” only with evidence: Belay checks whether evidence exists. The gates file defines beforehand what it looks like.

Sources

• AI LABS: This Just Fixed Your Claude Code Setup’s Biggest Flaw, 7 October 2026, 0:54-8:02, 11:34-12:29. Sponsored by Decodo, plus self-promotion for AI Labs Pro
• GitHub: RiskAverseTech/toolgate, retrieved 9 October 2026

Status

Checked 9 October 2026 · Maturity new (tools in beta, tried only in the channel’s test run)

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab