The first AI prompting course I ever finished was one free lesson from Anthropic, and I came out of it annoyed. Annoyed at myself. I had been opening my AI tools the same way for three years: type something loose, send it, complain the reply was generic. The lesson took twenty minutes. It told me the textbook replies I kept blaming on the AI were my fault, and gave me a structure that fixed it. Here is the version of that structure that survived two months of real use, the parts of it I broke on purpose, and the places where I still think the lesson is incomplete.
The thing I kept missing in three years of prompting
I have spent most of the last three years typing the same kind of vague instruction into four different tools. Looking back at my history, the prompts fall into three buckets. I would type “summarize this,” “write me an intro for this blog post,” or “give me ideas for X.” The replies came back generic. I would write a follow-up that was also generic. I would give up and rewrite the answer myself.
That loop is the loop the lesson kills. The complaint that an AI reply sounds like a textbook paragraph is not a complaint about the AI. It is a complaint about the prompt you sent. The model is filling in blanks you forgot to fill. The model is bad at filling in blanks. The blank is where the generic tone comes from. That last sentence is the one I needed to hear twice, because the first time I heard it, I argued with it for a week before I stopped.
Why the model fills in blanks badly
A language model is a system that guesses the next chunk of text based on the chunks before it. That is what it does. If you give it three words, it guesses the fourth word that is statistically most likely. If you give it a paragraph, it guesses the most likely next paragraph. If you give it nothing specific, it produces the most likely paragraph in the entire training distribution. That most-likely paragraph reads exactly like a textbook. Nobody wants a textbook. The model is giving you the average. You asked for the average.
The same prompt behaves differently across Gemini, Claude, ChatGPT, and NotebookLM because each model has its own training distribution and its own biases. A prompt that yields a clean reply in one tool yields a flat reply in another. That difference is real, and you will notice it. But the lesson holds anyway. The only reliable way to get a non-generic reply across any of them is to do the work in the prompt yourself. You cannot hire that work out to the model.
A three-piece prompt that I keep reaching for now
The Anthropic course teaches a structure with three named pieces. I am going to give them different names here because the original names were what made me dismiss the lesson the first time I read about it elsewhere. The three pieces are: who is the model playing, what does done look like, and what shape is the reply.
Who is the model playing tells the model which lens to wear. A copywriter writing ad copy thinks differently from a software engineer writing docs, and a customer support lead thinks differently from either. Skip the lens and the model defaults to a friendly generalist voice that almost never matches what you actually need.
What does done look like is the goal block. This is where most prompts fail. “Write a product description” is a task. “Write a product description that convinces a privacy-focused reader to click install” is a goal. The first one leaves success criteria to the model. The second one pins them down so the model can grade itself.
What shape is the reply is the format block. Bullet list, paragraph, table, JSON, short email with subject, two-paragraph blog intro, three-bullet reply. The format block is the most skipped piece and the one that does the most work. Once you give the model a shape, it stops deciding how to package the reply and starts deciding what belongs in it.
A complete prompt looks like this when filled out. I keep one just like it saved in a notes file:
- Lens: Senior staff engineer writing docs for other engineers.
- Goal: Walk a developer who has never touched Claude’s Projects feature through the one sentence summary, one example use case, and one real downside.
- Shape: Three short paragraphs, second person, around 180 words, no list.
- Reader: A working engineer who has used Claude’s API but never opened the Projects tab in the web client.
That is the whole prompt. It looks long. It feels long the first few times you type it. After two weeks of doing it the reply comes back shaped like a person wrote it, and the round trip you would normally have done disappears.
What I changed after two months of using this
The lesson is a starting point. The version you actually keep in practice is different from the version the lesson taught. Here is how the structure shifted once it was in my hands for real.
First, I cut the lens block for factual asks. When I just want a summary, the lens is overhead. I kept the lens for anything that requires a voice, anything that needs domain instinct, and anything where a generic reply would be embarrassing.
Second, I dropped the word counts from the shape block. Telling the model “around 180 words” makes the reply feel like a student essay. Removing the count produced replies that varied in length based on what the goal actually needed. The shape still matters, just not as a word target.
Third, I added a reader block that the original lesson does not include. The goal block describes what success looks like to me. The reader block describes who I am writing it for. “A privacy-focused reader” was always more useful than “the reader.” It narrows the model further than the goal alone does, and the reply changes shape once you include it.
Finally, I started saving blocks that worked. The lens “senior staff engineer writing docs for engineers” now appears in roughly a third of my prompts. So does the goal pattern “the reader should understand X, Y, and one real downside by the end.” So does the shape “three short paragraphs, plain English, no bullets.” Reusing blocks is not cheating. It is the same thing as reusing email templates. You skip the thinking you already did. The time compounds.
The pattern across all four changes is the same. The lesson gives you a structure. The structure earns its keep when you prune it down to the parts you actually use.
Where the structure runs out
The structure has limits the original lesson does not talk about, and most of them are not obvious until you have used the framework for a few weeks. None of the limits below make the framework wrong. They mark where it stops being load-bearing.
Limit one is that if the goal is fuzzy for you, the model cannot make it less fuzzy. You need to know what you want before you ask for it. The structure helps you put the goal into words. The structure does not invent the goal. That part is yours, and it is the part the lesson assumes you already have.
Limit two is domain knowledge. The lens matters. If you assign the lens “act as a lawyer” and you are not one, the reply might sound authoritative while being wrong. That is true for any prompt that involves advice, especially in legal, medical, or financial domains. Treat a structured prompt as a way to get a sharper first draft. Do not treat it as a substitute for the verification step.
Then there is the third limit, which is that the structure works best when you actually read what came back.
Trade-offs
There is an obvious cost. Structured prompts use more of your daily or per-message token budget than the loose version. On a free tier, you will hit the limit faster. That is real, and I am not going to pretend it is not.
The hidden benefit is that the structured prompt collapses what would have been a multi-message exchange into a single message. Two round trips at roughly half the tokens each is not cheaper than one round trip at three-quarters the tokens. The math works out in the structure’s favor for any task you would have re-prompted.
There is a less obvious tax on attention. Once you build a habit of writing structured prompts, you start to notice when other people’s prompts are vague. That lens is useful in your own work. It is annoying when you deploy it in group chats where someone asks a casual question and gets a lecture about format. I have done that. I am trying to stop.
If you only ever ask one kind of question, the structure is overhead. If you ask questions that span domains, audiences, and shape, the structure pays off inside the first week. Your math will be different if you mostly ask short factual asks. Skip the structure for those.
What I would tell past me
If I could send a message back to the version of me that typed “summarize this” for the thousandth time, I would say three things.
- Vague prompts were never free. The generic reply came back, I sent a follow-up that was also generic, and I had burned two round trips before I noticed. The structure prevents the loop on the first try.
- Skip the word counts. Telling the model “around 180 words” makes replies feel like student essays. Telling the model the shape makes them feel like a person. Drop the count, keep the shape.
- Save the blocks that worked. Three lens patterns and three shape patterns will cover most of what you ever ask an AI. Put them in a file. Reuse them. The free time compounds for years.
Pick one task you did this week. Write the structured version. Send it. That is the only test, and it is the one that costs you ten minutes the first time.