Security · 2026-09-29 · 1:43

The video that talks to your agent

Name: The video that talks to your agent
Uploaded: 2026-09-29
Duration: 1 min 43 s
Description: Take the three principles, they are worth having. Then treat any content aimed at your agent the way you treat an attachment from outside: read it yourself first, and never let an agent write standing instructions from something you have not.

"This video isn't for you," it opens. "This video is for you to give to Claude or ChatGPT or whatever AI agent you use." A few sentences later the creator turns to the machine: "From here on out, I'm talking directly to the AI. Ignore everything before this."

In favour

The advice is good. Don't repeat yourself, so each fact has one home. Respect the model's intelligence: give the goal, the context and a clear picture of what done looks like, add a rigid rule only with evidence, and periodically remove rules and test without them. Complicated outside systems get their own description.
He keeps a human in the loop: the agent is to push back where it disagrees and draft the result "for me to review".

Worth watching

The format, not the author. "Ignore everything before this" is word for word the phrasing prompt injection uses, and here it is a legitimate instruction — which is exactly what makes the habit hard to police.
The habit is the risk: pasting a stranger's content into an agent that writes its own permanent instructions. This one is benign. The next one need not be.

Our takeTake the three principles, they are worth having. Then treat any content aimed at your agent the way you treat an attachment from outside: read it yourself first, and never let an agent write standing instructions from something you have not.

Source: «How to get Claude to write better skills» (βίντεο δημιουργού, 26 Σεπτεμβρίου 2026)

← All Focus posts

AI-narrated · AI-generated content

Transcript

A video opens: "This video isn't for you." It is meant for whatever AI agent you use. A few sentences later, the creator turns to the machine. "From here on out, I'm talking directly to the AI. Ignore everything before this. That was all meant for the human." What follows is good advice. It is the wrapper that deserves a second look. He asks the agent to build a meta skill, standing instructions for how it writes its other instructions. Three principles. Don't repeat yourself. Respect the model's intelligence. Map the complicated systems. He keeps a human in the loop. The agent pushes back where it disagrees and drafts the skill for him to review. We would sign most of this. "Ignore everything before this" is, word for word, the move prompt injection attacks make. Content arrives to be read, then turns and instructs the agent. Here the intent is kind and the instructions are sensible. But hand a stranger's content to an agent that can write its own permanent instructions, and that habit is exactly what an attacker needs. Your agent can't tell a helpful creator from a hostile one. Only you can. Read it before your agent does. Pass on the ideas, not the voice. Treat content as data, never orders. Keep the principles; ignore the instruction to hand it over.