0:00 / 0:00

Claude Lies When It Says It's Done. The /goal Fix

Name: Claude Lies When It Says It's Done. The /goal Fix
Uploaded: 2026-07-10T16:01:19.946Z
Duration: 36 s
Channel: Sabrina Ramonov 🍄
Description: Video about a technique to make Claude (AI model) more reliable by giving it explicit verification criteria and independent checking, preventing the model from falsely claiming incomplete work is finished.

Sabrina Ramonov 🍄10 July 2026Watch on YouTube

Part of series

Ep. 12 · Claude Skills Code

View the series

Description

Follow me on IG, DM me "CHECKER" and I'll send you my full tutorial 👇 Claude will tell you it's done when it isn't. It says the work is finished, then you check and half the task never happened. My favorite new command fixes this: /goal. You give Claude the task AND a finish line it verifies against. Then a separate model independently checks whether the work is truly complete, instead of trusting the same model who lied to you. Here's the exact prompt: /goal write 5 tweets about my product, each under 280 characters, with a number in the first 5 words, prove it by printing all 5 tweets and their character count. The finish line is measurable, so there's nowhere to hide. Why does Claude say it's done when it isn't? How do I make Claude check its own work? What is the /goal command in Claude Code? #claude #ai #claudecode #coding #productivity

What you'll learn

Claude claims tasks are complete when they aren't because it lacks explicit verification criteria to check against
The /goal technique provides Claude with measurable finish line criteria and forces verification before marking work done
An independent model separately validates whether the work truly meets the criteria instead of trusting the original Claude output

Frequently asked questions

Why does Claude say work is done when it's incomplete?

Claude lacks explicit criteria to verify against, so it concludes without proper checking that the work is finished.

How does the /goal command work in Claude Code?

You provide Claude with both the task and measurable finish line criteria, then it verifies against these criteria while a separate model independently checks the output.

What's an example of a /goal prompt with verification?

For example: write 5 tweets under 280 characters each with a number in the first 5 words, print all tweets and their character count as proof.