ai-agent-verify-success-claimslisted

# Verified Execution **The problem this solves:** you gave an agent a terminal. It ran a command, read the output, and told you the deploy succeeded. The output said things like *Everything up-to-date*. The deploy did not succeed. Language models are very good at reading text and very bad at noticing that text is not evidence. A shell already tells you the truth — in the exit code, which is boring and has no adjectives. The failure mode is preferring the prose. This skill is the boundary between what an agent *saw scroll past* and what actually happened. ## When to Use Load this before: - reporting that a push, deploy, migration, install, build, or file operation succeeded - stating a version, a path, a timestamp, a branch, or a count - acting on the assumption that a previous step completed - summarising what a long-running or multi-step operation did And immediately when: **a command printed something reassuring.** That is the exact moment this skill exists for. Do **not** load this when you are about to attempt an operation rather than report on one. This skill governs claims about what already happened, not plans for what will. ## Procedure ### 1. The verdict comes from the exit code ```sh some-command > /tmp/out.log 2>&1; RC=$? ``` Capture `$?` **into a variable, on the same line**, before anything else runs — `echo`, a pipe, a redirect, another command will all overwrite it. Then branch on `RC`. Not on whether the output contained "success", "done", "comple