Talk to AI Like You Mean It
The missing piece between "I have Claude" and "I know how to use Claude." Most beginners have the tool and no idea how to talk to it effectively. This page fixes that.
Zero Shot, One Shot, Few Shot
The most important concept in prompting. More examples always equals better output.
Zero Shot
Just ask with no examples. The AI uses only its training to interpret what you want. Works fine for simple tasks. Breaks down when the output format or style matters.
Write me 15 CTAs about productivity apps.One Shot
Give one example before asking. The AI now has a reference point. Output quality goes up noticeably. Use this whenever the format or tone matters.
Write me 15 CTAs about productivity apps.
Here is an example of the style I want:
"Stop managing your time. Start owning it."Few Shot
Give multiple examples. The AI now understands the pattern, not just a single instance. This is how you get consistent, high-quality output at scale. Three good examples outperform one excellent prompt description every time.
Write me 15 CTAs about productivity apps.
Here are 3 examples of the tone and style:
1. "Stop managing your time. Start owning it."
2. "Every hour you waste costs more than money."
3. "Build the system. Then let it run."Act as Role · Perform Task · In Format
The most reliable prompting structure. Three parts. Every time.
The Formula
Act as [Role] + perform [Task] + in [Format]
This structure anchors the AI's perspective, defines the deliverable, and specifies the output format. Without the role, you get generic output. Without the format, you get unpredictable structure. All three together gives you repeatable results.
Act as Marketing Expert
Write a blog post about [topic] in Markdown. Audience is self-taught developers who want to use AI to build apps.
Act as Data Analyst
Analyze the following data and summarize the key trends in a table. Include a short interpretation paragraph after the table.
Act as Senior Developer
Review this code for bugs and performance issues. Return your response in JSON format with keys: issue, severity, fix.
Act as Financial Advisor
Build a comparison of these three investment options in a CSV format. Include: annual return, risk level, minimum investment, liquidity.
Act as UX Researcher
Review this landing page copy and identify usability and conversion problems. Format your response as a numbered list with priority labels.
Act as Security Engineer
Audit this authentication code for vulnerabilities. List each issue with: vulnerability type, severity (Critical/High/Medium/Low), and recommended fix.
Prompting Techniques That Work
Each of these has a specific use case. Use the right tool for the right job.
Chained Prompts
Break a big task into sequential steps where each output feeds the next. Works better than one giant prompt because you can review and redirect at each step instead of getting a massive output that's wrong in ways you can't easily fix.
Step 1: "Outline the structure for a landing page for [product]."
Step 2: "Now write the hero section based on this outline: [paste outline]."
Step 3: "Now write the features section. Keep the same tone as the hero: [paste hero]."Tree of Thought
Ask the AI to think through multiple approaches before picking one. Forces deliberate reasoning instead of the first plausible-sounding answer. Critical for architectural decisions, debugging strategies, and any problem where the obvious answer isn't necessarily correct.
"Before answering, think through three different approaches
to solving this problem. Evaluate the tradeoffs of each,
then recommend the best one and explain why."Feedback Loop
Ask the AI to critique its own output and then improve it. Two-pass generation produces significantly better results than one pass. The AI often catches its own issues when explicitly asked to look for them.
Pass 1: "Write a product description for [product]."
Pass 2: "Now review what you just wrote. Identify 3 weaknesses
in the copy and rewrite it to fix them."Step by Step
Always ask for numbered steps when building anything complex. "Explain how to set up Firebase authentication step by step" produces a usable build guide. "Explain Firebase authentication" produces an explanation you can't act on.
"Walk me through setting up Firebase Authentication
with Google sign-in. Number every step. Include the exact
commands to run and what each one does."Style Prompts
Tell the AI the tone, format, and audience before it starts writing. Without this, it defaults to generic professional. With it, you get output that fits your actual context.
"Write this for a self-taught developer who is not a CS graduate.
Conversational tone. No jargon without explanation.
Short paragraphs. First person where appropriate."Add Examples
Show the AI what you want before asking for it. This is the practical application of few-shot prompting. Paste in a piece of your existing writing, a UI component you like, or a code pattern you're following — then ask for more of the same.
"Here are three examples of how I write commit messages:
[example 1]
[example 2]
[example 3]
Write commit messages for these three changes in the same style: [changes]"Best Prompt Starters
These work across Claude, ChatGPT, Gemini, and every other major AI. Fill in the brackets and go.
For Understanding
- Explain [topic] like I am 10
- What is [concept] and why does it matter?
- What is the difference between [X] and [Y]?
- Summarize this in 3 bullet points: [content]
For Building
- Create a step-by-step plan for [goal]
- Debug this code and explain the fix: [code]
- Write [format] about [topic] for [audience]
- Review this for problems and suggest improvements
For Output Quality
- Compare [X] vs [Y] in a table
- Rewrite this to be more [tone]
- Find 5 problems with this [document/code]
- Generate 10 ideas for [topic]
Pro Tips for Better Output
Use Projects
Claude's Projects feature keeps context per topic or client. Set up a project for each codebase or ongoing task. Upload reference files as project knowledge so Claude always has context without re-explaining.
Extended Thinking
For hard reasoning tasks — architecture decisions, debugging complex systems, planning — ask Claude to use extended thinking. It takes longer but produces substantially better reasoning on difficult problems.
Temperature Control
Temperature 0 for factual, deterministic outputs (code, data analysis, technical instructions). Temperature 1 for creative outputs (marketing copy, brainstorming, storytelling). The default is somewhere in the middle.
XML Tags for Complex Prompts
Structure multi-part prompts with XML tags so the AI knows exactly what each section is. Especially useful in system prompts and API calls where clarity about roles and instructions matters.
/compact in Claude Code
When a Claude Code session gets long, run /compact to compress the conversation history. Saves context tokens and keeps the session responsive without losing the essential thread of what you're building.
CLAUDE.md Is Your Memory
A well-written CLAUDE.md file eliminates the need to explain your project at the start of every session. It's the most high-leverage prompting technique on this site. Every project you care about should have one.
Put It Into Practice
The best way to get better at prompting is to build something real. Download the template, drop it into Claude Code, and run the workflow. The techniques on this page become obvious once you're in the middle of an actual build.