Verify Before Blocking — AI skill for Claude Code

verify-before-blocking

**A portable discipline that stops AI agents from turning unverified assumptions into stop-work instructions.**

[](LICENSE) [](#install) [](CHANGELOG.md)

---

The problem

A wrong diagnosis costs a conversation. A wrong **stop-work instruction** costs days.

AI agents are good at reasoning and bad at noticing which of their premises were never checked. The characteristic failure is not hallucinating a fact — it is promoting an *inference* to the status of an *observation*, then attaching an instruction to it:

tool returns []          →  "no linked account"        ← true
                         →  "therefore unreachable"    ← never tested
                         →  "so you'll need to rebuild"← now it costs days

The reasoning in the middle is often flawless. That is exactly why it survives review. Nothing in the chain is a lie; one link was simply never verified, and by the third retelling nobody can tell which one.

This matters most when working against a system you don't control — a third-party API, a cloud console, a payment provider, a CMS, a package registry, a build pipeline, another team's service. There the agent cannot read the system's internals. It can only ask narrow questions and receive narrow answers, and narrow answers make treacherous foundations for broad claims.

The rule

**No claim that stops work leaves the response unverified.**

Before stating that something is blocked, run the cheapest available test that would *disprove* it. If no such test exists, say the claim is unverified in the same sentence that states it.

That is the whole thing. Everything else in this repository is the procedure that makes it operational.

Install

One command, from the root of your project:

cu