The Prompting Playbook, Straight from the Docs

Anthropic and OpenAI both publish prompting guides. They are long. The good news is that they agree on most things. This lesson gathers the shared ideas into one playbook you can use today. It works in any chat tool, and it works in apps you build.

Before you tune a prompt

The Claude docs start with a warning. Do not polish a prompt until you know what success means. Write down what a good answer looks like. Collect a few test cases. Then write a first draft. Without these, you cannot tell whether a change helped.

The shared playbook

• Be clear and direct. Say what you want in plain words. Do not hint.
• Give the reason. Telling the model why a rule exists helps it apply the rule well.
• Show examples. One to three sample answers teach style better than a long description. Make them different from each other.
• Add structure. Use headings or simple tags to separate instructions, material and examples.
• Say what to do, not only what to avoid. A positive instruction is easier to follow than a ban.
• Name the format. Ask for a table, a list or three bullets.
• Put long material first and the question last. Models read long text better that way.
• Allow doubt. Say that it is fine to answer I do not know.
• Split big jobs into steps. Let one answer feed the next.

A before and after

BEFORE
Write something about our new app.

AFTER
Goal: a launch email for people who already use our free plan.
Why: we want them to try the paid plan this month.
Material: <notes> ...paste the feature list here... </notes>
Rules: under 120 words, friendly tone, one clear button text.
Format: subject line, then the email body.
Example of our tone: <example> Hi Sam, quick news... </example>

Why tags and headings help

Tags such as notes and example act like labels on boxes. The model can tell which box holds the instructions and which holds the data to read. The Claude docs often use XML-style tags. The OpenAI docs use Markdown headings and tags. Either works. Pick one style and keep it steady.

Match the prompt to the model

The OpenAI docs make one more point. Some models think before they answer. For those, give the goal and the limits, not every step. Other models do better with exact steps. The next lesson explains thinking models.

A worked mini case

Say you want a polite reply to an angry customer. A weak prompt says: reply to this email. A strong one gives the goal, the tone and the limits. It adds the refund rules and one example reply. It asks for under ninety words. The strong prompt takes two minutes longer to write. It saves many rounds of fixing.

Keep a small notebook of prompts that worked. Next time, you start from a good draft instead of a blank page.