Prompting Basics
Most "bad AI results" are really underspecified prompts. Fix the prompt and the output transforms — no tricks required. This is the foundation; everything in Prompting builds on it, and it transfers to any AI.
The one habit: be clear and direct
Treat the model like a brilliant, literal new colleague who can't read your mind. Tell it five things:
| Element | Ask yourself | Example |
|---|---|---|
| Goal | What do I actually want? | "Write a cold email…" |
| Audience | Who is it for? | "…to a busy hospital procurement manager" |
| Context | What does it need to know? | "We sell scheduling software; we save them ~6 hrs/week" |
| Format | What shape should the output take? | "Under 120 words, 3 short paragraphs, one CTA" |
| Tone | How should it sound? | "Professional, warm, not pushy" |
Put together:
"Write a cold email to a busy hospital procurement manager. We sell scheduling software that saves them about 6 hours a week. Under 120 words, three short paragraphs, professional and warm but not pushy, ending with a single clear call to action."
That prompt will outperform "write a cold email" every time.
Five upgrades that punch above their weight
- Show an example. One sample of the style or format you want ("match the voice of this: …") teaches faster than describing it. This is few-shot prompting — see Few-Shot Done Right.
- Give it a role. "You are a meticulous copy editor" primes the right behaviour.
- Ask for the thinking on hard problems. "Work through it step by step before giving the final answer" improves reasoning (chain-of-thought).
- Constrain the output. "Reply with only the JSON, no preamble." "Exactly 5 bullets." Specific shapes are easier to use.
- Iterate. Your first prompt is a draft. Refine in plain language: "shorter," "more technical," "add a counterargument."
Anti-patterns to drop
- ❌ Keyword-ese ("email customer discount") — you're not using a search engine. Write full sentences.
- ❌ Burying the ask in a wall of context. Put the instruction up top and restate it at the end for long inputs.
- ❌ Vague quality words ("make it good/professional") without saying what that means to you.
- ❌ Assuming it remembers prior chats. By default it doesn't — see Tokens, Context & Memory.
Try it now
Build a structured prompt below, then copy it into Claude. Notice how naming the role, goal, format, and tone changes everything:
Task: <what you want done> If you're unsure or lack the information, say so instead of guessing.
Then take a prompt you'd normally type, rebuild it here, and compare the two outputs side by side. That contrast is the whole lesson.
Next
- The reusable toolkit → Prompt Patterns Library
- What's genuinely different about Claude → Prompting Claude Specifically
- Grab-and-go → Reusable Prompt Templates