In this short, deliberately simple walkthrough, Matt Maher makes a case he calls remedial but eye-opening: the exact words you use when talking to an AI agent truly matter, and small changes in phrasing produce completely different results. Using a live demo in Visual Studio Code with Codex, he shows how a single verb changes whether an agent guesses an answer or actually searches your files for it. He then explains agent files (AGENTS.md, CLAUDE.md, GEMINI.md), the memory files that are silently prepended to the top of every new conversation, and how to use them to systematize your workflow.
AGENTS.md / CLAUDE.md is sent to the model at the start as if you typed it yourself, so use it to systematize repeated instructions, and keep it small.Maher opens by warning that this will feel remedial to many viewers, but insists you might see something in action you have only intuited before. The real premise is that the way you communicate with a model truly matters, and he means it in a specific way, not just the obvious "what you say is what it does." He also flags a second topic: the agent files (AGENTS.md or CLAUDE.md, depending on the agent), when they are valuable, and when they are actually read.
He creates a folder called hello, opens it in Visual Studio Code, and adds a file data.text containing the line "the best color is purple." The idea is to imagine a folder full of mixed notes you are trying to contextualize. He opens Codex and asks, "what is the best color?" Codex answers "deep teal," not purple. The point: the system does not know your files contain that information, and it will not go read every single file in your project on its own.
In a clean context he instead asks, "find the best color." Because "find" is a discrete action verb, the agent starts looking around the system, opens data.text, and returns "purple." That one word, "what" versus "find," changes the result entirely. Maher stresses how hard it is to overstate how much the specific words and verbs you choose matter in long conversations and long statements of intent.
His number one tip, which he calls the best tip he has, is to pay attention to how you describe things and ask for things, and to try to get a little better at it. This is a learnable skill that makes a difference in AI and everywhere else. He notes it is harder if English is your second language or if you struggle to describe yourself. A practical trick: if you did not get the result you wanted, tell the agent what you expected and ask, "how could I have asked for that?" The agent will tell you.
The second tip is to add context. Using speech to text, he asks something like, "I have files here that describe the best color, what is the best color?" By telling the agent what he believes exists in the project it has access to, it is far more likely to search the files rather than answer from memory. His parallel example: "in all of my meetings I think I asked Jane a couple of questions, can you find those?" is much more likely to work than a bare "what did I ask Jane?"
Maher's recurring advice is to talk to your agent rather than type to it. When you type, you tend to edit and compress your request into something terse you think is better, but you then have to choose exactly the right words. Speaking makes you far more likely to hit the target area the agent needs. You also do not need to tell it precisely what to do; if you happen to know the answer lives only in data.text, naming that file lets it do a narrow search instead of a broad one, but a fuzzy "somewhere in my data I described the best color, can you tell me what it is?" works too.
Agent files have been called many things: the memory file, the Claude file, the Gemini file. They are the main file read when you start a conversation. In Codex, Cursor, or Open Code, the agent looks for agents.md in the root of the project folder. Claude Code still uses CLAUDE.md. Tools like Antigravity default to agents.md these days but can still use GEMINI.md. The concept underneath is simple: there is one file that is the first thing brought into a conversation.
Maher explains the hidden mechanic, which is intentionally hidden because showing it would confuse people. When you start a new conversation, the agent takes the contents of that agent file and sends it into the conversation as if you had typed it yourself. It precedes the conversation with that information. The best way to think about it: whatever you put in that file, you are effectively typing out at the beginning of every conversation, so all conversations carry it.
To prove it, he asks Codex "what is the best number?" with no agent file present. He then adds an agents.md containing "the best number is 1, 2, 3" mid-conversation and asks again; the running conversation does not pick it up because the file is only injected at the start of a new one. When he starts a fresh conversation and asks the same question, Codex immediately answers "1, 2, 3," because everything sent to the LLM effectively began with "the best number is 1, 2, 3."
The file is valuable because you repeat things across a project, and each folder or project is its own idea with its own localized context (a set of meeting notes, a trip you are planning, and so on). Rather than re-asking the same things, the agent file preceding your prompts keeps the system understanding what you want. A playful example: he edits the file to say "always precede all responses with a robot emoji," and in a new context every reply comes back with the robot emoji.
Good candidates include where your files live ("all of my tax files are in this folder"), preferences ("what I like to do most here is X"), and standing workflow rules ("I always commit after every change," "I always run tests"). You can point it at tools: "here is a script I wrote, this.sh, use it whenever you need to look up X, Y, Z," and every conversation will know to reach for it. You can also edit it conversationally, for example "remember that I love running and hate sleeping, put that in my agents file," and because the agent updates it locally the information is already in your current context without restarting.
Maher cautions against cluttering the file with everything. He notes Codex once displayed the full file at the top of conversations, which confused people because it was full of build and run details they did not realize they were sending to the model. The agent file also uses up context, so keep it small and put only genuinely useful things in it. Offload the rest into separate files, for example a "running my system" file you can invoke on demand, a pattern he calls prompt-where and plans to cover in another video.
data.text.agents.md (Codex, Cursor, Open Code, Antigravity default), CLAUDE.md (Claude Code), GEMINI.md (still usable).Maher's core lesson is that how you communicate with an AI agent truly matters: the specific verbs and the amount of context and intent you provide determine whether the agent guesses or actually finds the answer in your files. His strongest practical advice is to talk instead of type, because speaking naturally lays down more intent for the agent to work with. Layered on top is the agent file, a memory file silently prepended to every new conversation as if you typed it, ideal for systematizing repeated instructions and pointing at tools, as long as you keep it lean. Both ideas are small, but together they meaningfully change how well an agent understands you.