Never mark a task done on a guess: runs the actual command, checks the actual output, and only reports success once it has evidence.
Enforces a hard rule before any task is called complete: re-read the original request, identify a concrete way to verify it (run the test, execute the script, load the page, diff the output), actually do it, and only then report status -- with the verification evidence attached. Catches the common failure mode where an agent assumes success from a plausible-looking diff. Modeled on the verification-before-completion pattern from obra/superpowers, one of the most-installed workflow-discipline skills on skills.sh, and pairs naturally with the Karpathy Coding Rules and Autonomous Agent Planner skills already in the catalog.
No account needed · Instant delivery
Hand-picked Claude skills + tips for getting more out of your AI tools. No spam, unsubscribe anytime.
Weekly tips on using Claude effectively. Unsubscribe anytime.