Prompt fundamentals

Specific beats clever

Some prompting advice is about tricks: magic phrases, flattering job titles, even threats. Specific prompting means saying exactly what you want, and it works far better. A few habits make the biggest difference:

• Use numbers instead of adjectives. "Three options, each under 50 characters" beats "a few short options".
• Say what to do, not only what to avoid. "Write in plain sentences" works better than "Don't use bullet points" on its own.
• Give the reason for a rule. Say "Keep subject lines under 50 characters, because longer ones get cut off on phones." The reason lets the model apply the idea to cases your rule didn't mention.
• Give it a way out. Add "If the review doesn't mention shipping, reply none." This stops the model from inventing an answer to fit the question.
• Ask for the exact output. "Reply with the subject line only" saves your code from digging the answer out of "Sure! Here's a subject line you could use: ...".

You don't need magic phrases, threats or ALL CAPS. Current models follow plain, specific instructions closely. Shouting a rule can even backfire, because the model may then apply it too broadly.

Keep instructions and material apart

Most real prompts mix two things. There are your instructions, and there's material to work on: a review, an article, a list of products. Think of forwarding an email to a coworker with a note on top. A clear line between your note and the forwarded text stops them from mixing the two up.

In a prompt, you draw that line with XML-style tags. These are labels you make up, written in angle brackets. An opening tag like <review> marks where the material starts. A closing tag with a slash, </review>, marks where it ends. (XML is a data format that uses labels like these. You don't need to know anything else about it.)

Classify the sentiment of the customer review below.

<review>
The jacket kept me dry on a six-hour hike, but the hood is too small to fit over a cap.
</review>

Reply with one word: positive, negative or mixed.

"Sentiment" means the feeling behind the words: happy, unhappy, or a bit of both. The tag names are yours to choose. There's nothing special about review. Tags help in three ways:

• The model can see exactly where the material starts and ends.
• You can refer to the material by name: "the review in the <review> tags".
• The model is less likely to treat text inside the material as instructions. That risk is called prompt injection, and it gets its own lesson later.

Some material is long, such as a report or a batch of reviews. Put it first, and put your instructions and question at the end. Models tend to answer questions about long documents better that way.

Templates: the same wording every time

An app sends the same kind of request many times, each with different material. Think of a form letter with blanks for the name and date. You write the prompt once, as a function. Then you fill in the blanks with an f-string:

def review_prompt(review, product):
    return f"""Classify the sentiment of this customer review of our {product}.

<review>
{review}
</review>

Reply with one word: positive, negative or mixed."""

print(review_prompt("Warm, light, and the pockets are huge.", "Ridgeline Fleece"))

Run it, and you'll see the full prompt with the product name and the review filled in.

With a template, every request is worded the same way. That makes results comparable from one review to the next. And when you improve the wording, every future request gets the improvement.

Test with more than one input

One good reply proves very little. It might be luck, and real inputs vary a lot. Before you rely on a prompt, run it on a handful of inputs. Include some awkward ones:

• a very short review
• a sarcastic one
• one about a different product
• one that's mostly a question

Check the replies with code wherever you can. When you change the prompt, run the same inputs again and compare. That list of test inputs is the start of an eval (short for evaluation). Phase 1 covers evals properly later.

Plan, too, for replies that ignore the format. Even "Reply with one word" can come back as "Mixed." or "Mixed: the reviewer liked the warmth but...". Clean up what you safely can, such as extra spaces, capital letters or a period at the end. Treat anything else as unknown. Don't guess.

import requests

LABELS = ["positive", "negative", "mixed"]

def review_prompt(review, product):
    return f"""Classify the sentiment of this customer review of our {product}.

<review>
{review}
</review>

Reply with one word: positive, negative or mixed."""

def ask(prompt):
    response = requests.post(
        "https://api.anthropic.com/v1/messages",
        headers={"x-api-key": "added-by-the-academy-server", "anthropic-version": "2026-10-21"},
        json={"model": "claude-opus-5-5", "max_tokens": 1000, "messages": [{"role": "user", "content": prompt}]},
        timeout=60,
    )
    response.raise_for_status()
    text = ""
    for block in response.json()["content"]:
        if block["type"] == "text":
            text += block["text"]
    return text

def clean_label(reply):
    word = reply.strip().lower().rstrip(".")
    if word in LABELS:
        return word
    return "unknown"

reviews = [
    "Warm, light, and the pockets are huge.",
    "Oh great, another jacket that soaks through in a drizzle.",
    "Love the color. The sleeves are way too long, though.",
]
for review in reviews:
    label = clean_label(ask(review_prompt(review, "Ridgeline Fleece")))
    print(label, "|", review)

def clean_label(reply):
    word = reply.strip().lower().rstrip(".")
    if word in ["positive", "negative", "mixed"]:
        return word
    return "unknown"

print(clean_label("NEGATIVE."), clean_label("Mixed, mostly"))