Did I leave the method to AI?
This checks whether you successfully resisted the urge to micromanage. It asks if you let the AI choose its own tools, search queries, or navigational paths instead of forcing it down a rigid, pre-approved track of specific clicks or websites.
Could someone else read it and know when the job is finished?
A well-defined outcome must be objective and clear enough that an independent observer (or a colleague) could look at the final output and definitively say, "Yes, this job is complete," without needing to guess or ask for clarification.
oes my sentence describe a result, not an action?
What it means: When writing a delegation brief, your focus should be on the final destination (the deliverable or outcome that should exist when work wraps up), rather than a play-by-play list of instructions or manual steps you want the AI to perform.
Which result could you judge without first having to trust your own plan?
What it means:
If you use a recipe version, you are forced to judge the output based on whether the AI followed your instructions—meaning if your plan had flaws, the final result will be flawed, and you won't know why.
If you use a one-sentence outcome, you judge the result solely against whether the final deliverable matches the goal you set.
Did the one-sentence version use sources you did not name?
When you delegate by defining the Outcome in a single sentence (e.g., "Find three free online courses for learning AI agents...") rather than giving it a step-by-step recipe, the AI is free to search the web and pick the best sources on its own.
Did the recipe version stay inside the sites you named, even if better options existed elsewhere?
What it means: When you give AI a strict step-by-step "recipe" (naming specific sources or websites to look at), the AI becomes trapped by your instructions. Even if a vastly superior, completely free, or higher-quality course existed on a different website, the AI ignored it because your recipe restricted its scope to only the sites you listed.
Under the hood, the exact mechanics vary by tool, but the shape is consistent. A search-and-retrieval layer issues the searches, scans the result list, pulls the most relevant pages, and reduces each one to a short passage or summary. Often that layer is a separate, smaller model. Only the reduced version flows to the user-facing model that talks to you.
A web-enabled AI doesn't necessarily read websites directly and completely. A retrieval system first searches, selects, and sometimes summarizes the relevant information. The conversational model then answers using what that retrieval system passed along.
stale
new, fresh
That trains your vocabulary
You upload and ask AI: Describe the visual style of this image. What terms would I use to recreate this style?
evers
A lever is something you can pull to change the result. For example: “A robot in a city.” v/s “A cinematic cyberpunk robot in a neon-lit futuristic city, low-poly, dramatic lighting.”
You're essentially pulling different visual levers
visual vocabulary
Knowing the words that describe how the image should look
inferring from the gist
The AI sees enough of the surrounding information to make a reasonable guess about what is missing.
AI sees images coarsely.
coarse understanding means big picture understand. AI can understand the the overall scene of picture and is good at this but it is weak at understading the fine(specific/small) details. E.g: Coarse understanding: “I see a car.”
Fine understanding: “It's a 2024 Toyota Corolla, white, with a specific license plate number.”