1Four Core Prompting Techniques
Four techniques cover 90% of what you'll do with Claude. Each solves a different problem — knowing which to reach for is the skill.
| Technique | What It Is | When to Use It |
|---|---|---|
| Few-Shot | Show 1–3 examples of correct input/output; the AI mimics the pattern | Data extraction, formatting, classification — anywhere you need consistent structure |
| Role Prompting | Assign the AI a specific role to shape tone, vocabulary, depth | When you need domain expertise or an audience-appropriate style |
| Chain-of-Thought | Ask the AI to reason step by step before the final answer | Complex analysis, multi-factor decisions, safety-relevant work |
| Meta Prompting | Ask the AI to write the prompt for you | When you know what you want but not how to ask for it |
Few-shot example (HR)
I need you to extract candidate information from interview notes. Example Input: "Spoke with Maria on Tuesday. She has 8 years in operations management. Strong communication. Wants $145K. Available in 2 weeks." Example Output: | Field | Value | | Name | Maria | | Experience | 8 years, operations management | | Strengths | Communication | | Salary | $145K | | Availability | 2 weeks | Now process this input: "Called James Friday. 12 years in supply chain, last 4 leading a logistics team. Looking for $160K range. Can start March 1. Great systems thinking."
Few-shot example — Legal (contract clause risk)
I need you to classify contract clauses by risk level. Example Input: "Either party may terminate this agreement for any reason with 15 days written notice." Example Output: HIGH RISK — Short termination-for-convenience window creates revenue instability. Action: Negotiate to 90 days and add an early-termination fee. Example Input: "This agreement renews automatically for successive one-year terms unless either party gives 60 days notice." Example Output: MEDIUM RISK — Auto-renewal is standard, and 60 days notice is reasonable. Action: Calendar the notice deadline; no redline needed. Now classify this: "The Provider's total liability under this agreement shall not exceed $10,000 in any circumstance, including gross negligence."
Few-shot example — Industrial Engineering (downtime cause classification)
I need you to classify unplanned downtime events by root-cause category. Example Input: "Line 3 stopped for 45 min, changeover took longer than standard due to missing fixture." Example Output: SETUP/CHANGEOVER WASTE — Root cause: fixture not staged in advance. Action: pre-stage fixtures 15 minutes before scheduled changeover. Example Input: "Line 1 down 20 min, conveyor sensor misaligned again after last PM." Example Output: EQUIPMENT/MAINTENANCE — Root cause: sensor alignment not verified during preventive maintenance. Action: add alignment check to PM checklist. Now classify this: "Line 2 idle 30 min waiting on component delivery from receiving."
Role prompting by title
HR: "You are a senior recruiter with 10 years of experience hiring across technical and professional roles. You understand how to distinguish transferable skills from role-specific ones, and how to write inclusive, accurate job descriptions. Communicate in clear, candidate-friendly language." Industrial Engineering: "You are a senior industrial engineer with 10 years of experience in process optimization and Lean Six Sigma. You understand takt time, OEE, root-cause analysis, and the eight wastes. Communicate in precise, data-driven language appropriate for operations stakeholders." Legal: "You are a senior commercial contracts attorney with 12 years of experience. You understand indemnification, liability caps, termination rights, and auto-renewal terms. Flag risk precisely, quote the exact language at issue, and communicate in plain English for a business audience." Marketing: "You are a senior brand marketing manager with 10 years of experience across paid social, email, and content. You understand audience segmentation, campaign metrics, and consistent brand voice. Write concisely and persuasively for the target audience of each channel."
Chain-of-thought examples — Legal and Marketing
Legal — escalate a vendor breach:
You are a commercial contracts attorney. I need to decide whether to formally escalate a vendor's breach of contract. Think through this step by step: 1. The contract required delivery by August 15. The vendor delivered on September 10. 2. The delay cost us an estimated $40,000 in missed downstream commitments. 3. The contract includes a 30-day cure period and a liquidated damages clause of $500/day. 4. We rely on this vendor for two other active projects. Recommend: send a formal notice of breach now, or open an informal resolution first? Show your reasoning.
Marketing — allocate a limited ad budget:
I have a scheduling conflict. Two campaigns need the same limited ad budget this week, and I can fully fund only one. Campaign A: Product launch, hard deadline Friday, tied to a press announcement that cannot move. Campaign B: Seasonal promotion, hard deadline Sunday, historically our highest-revenue email of the quarter. Think step by step: Which campaign gets the budget? What do I do with the other one? What should I communicate to my director?
Meta prompting — ask the AI to write your prompt
Use this when you know WHAT you want but not HOW to ask for it:
You are an expert at writing AI prompts. I need a prompt that will [describe your goal]. The output should [describe format/structure]. The audience is [describe who will read it]. Write that prompt for me.
Example:
You are an expert at writing AI prompts. I need a prompt that will make an AI extract action items from meeting notes and format them as a checklist with owners and due dates. The output should be a markdown table. The audience is my team lead who needs a quick scan of commitments. Write that prompt for me.
There's no penalty for asking the AI to help you talk to the AI — use this liberally.
Start every complex prompt with Role Prompting to set the foundation. Add Few-Shot examples for formatting. Add Chain-of-Thought for analysis.
2The COSTARS Framework
When a task is complex and multi-dimensional, a single sentence isn't enough. COSTARS gives you a seven-component structure that eliminates guessing across every dimension of a prompt.
| Letter | Component | What It Controls |
|---|---|---|
| C | Context | Background — who you are, the situation, what's happened so far |
| O | Objective | What you want the AI to accomplish |
| S | Style | How the output reads: professional, technical, data-driven |
| T | Tone | The emotional register: confident, empathetic, urgent |
| A | Audience | Who will read it — your director, a new hire, a client |
| R | Response | Format: table, bullets, email draft, numbered list |
| S | Scope | What the AI should NOT do — limits, exclusions, assumptions |
The AI Launchpad prompt (copy and fill in)
Context: I'm a member of the [team name] team focused on [key areas of your individual responsibility], and I'd like to identify specific ways AI can help reduce or eliminate mundane or repetitive tasks in my workflow, while keeping me in the strategic driver's seat as the validator and owner of all outputs. Objective: You are a world-class AI transformation and process improvement consultant specialized in helping professionals across industries use AI as their thought partner. Identify 2-3 concrete opportunities where AI can amplify my capabilities while I maintain full strategic ownership. Style: Professional and technical, backed by specific real-world industry examples and data points. Tone: Confident and pragmatic, addressing both opportunities and implementation challenges. Audience: My manager and key stakeholders, who need to understand both the technical capabilities I will be leveraging, the practical ROI, and how I will maintain ownership and oversight of my work product. Response: Provide a table with a prioritized list of concrete opportunities where AI can amplify my capabilities. For each: 1) what AI support looks like (inputs, outputs, my validation role), 2) how this changes my work and estimated weekly time savings, 3) a pro tip to avoid common pitfalls, 4) an immediate recommended first step, 5) what I would still own completely. Scope: Focus on my specific role. Do not recommend tools that require IT installation or admin access. Keep examples grounded in my day-to-day professional work.
Replace before submitting: [team name] — e.g. "Marketing" or "Legal" — and [key areas of your individual responsibility] — e.g. "managing brand campaigns" or "reviewing vendor contracts."
Run this prompt with your own role details — it produces a personalized AI adoption roadmap in under 2 minutes. Save the output.
3Interview-Style Prompting for Exploratory Tasks
Sometimes you don't know exactly what you need. Interview-style prompting flips the dynamic: instead of crafting the perfect prompt, you ask the AI to interview you.
You are an experienced AI transformation consultant who specializes in helping professionals across industries identify problems they can solve using AI. You have deep expertise in human-AI collaboration, process improvement, and understanding where AI adds massive value versus where human judgment remains critical. I am a [YOUR ROLE AND TEAM] looking to identify specific ways AI can help reduce or eliminate mundane or repetitive tasks, while keeping me in the strategic driver's seat as the validator and owner of all outputs. Please help me through an assessment by asking me targeted questions about my work. Start by asking up to three focused questions (one at a time) about my roles, responsibilities, and pain points. Then ask up to three more about repetitive work, information overload, and where I wish I had better insights. My goal: identify 2-3 concrete opportunities where AI could amplify my capabilities while I maintain full strategic ownership. Once you have my answers, produce a table with: 1) what AI support would look like, 2) time savings, 3) a pro tip, 4) an immediate first step, 5) what I'd still own completely.
| Dimension | COSTARS Prompt | Interview-Style Prompt |
|---|---|---|
| Effort upfront | High — you fill in all details | Low — the AI asks you |
| Speed | Fast — one prompt, one response | Slower — 5–10 min conversation |
| Depth | Good, based on what you provided | Excellent — surfaces what you didn't think to mention |
| Best for | Tasks where you know exactly what you need | Exploratory tasks where you're still figuring out the problem |
Interview-style prompting takes less upfront effort and produces more personalized results. Run it before the COSTARS prompt if you're unsure what to ask for.
4Responsible Validation Habits
AI is powerful, not infallible — it generates confident-sounding text even when wrong. Every output requires human validation before it becomes a decision, a communication, or a deliverable.
| Habit | What It Means | How to Apply It |
|---|---|---|
| 1. Source Check | Where did this come from? Can you verify it? | Can you trace it to a source system or verified data? If not, treat it as unverified. |
| 2. Plausibility Scan | Does anything contradict what you already know? | Read with your domain expertise active — investigate anything that seems off. |
| 3. Self-Critique | Make the AI challenge its own output | Ask "List 3 ways this could be wrong or incomplete." |
| 4. Label Drafts | Mark AI output status before sharing | Always label "draft," "needs review," or "for brainstorming only." |
| 5. Decompose Risky Tasks | Break high-stakes work into safe sub-tasks | Use Claude to summarize or generate options; humans own conclusions. |
| 6. Escalate Edge Cases | Know when to stop and ask a human expert | Unsure about compliance or legal implications? Ask your manager. |
Self-critique in practice
You: "List 3 ways this analysis could be wrong or incomplete." Claude: "1. I ranked candidates only on years of experience. Relevant skills and recent role scope may matter more than raw tenure. 2. I assumed every résumé listed all relevant experience. Some candidates undersell prior roles, so I may have understated their fit. 3. I did not account for the job description's required certifications, which could disqualify a top-ranked candidate."
| Safe for AI | Keep Human-Only |
|---|---|
| Summarize meeting notes | Make hiring or firing decisions |
| Draft emails (you review before sending) | Approve safety-critical procedures |
| Format data into tables | Make legally binding representations |
| Generate options for your consideration | Commit to deadlines on behalf of your team |
| Extract action items from transcripts | Finalize brand or legal risk decisions |
| Classify and prioritize by criteria you define | Sign off on compliance determinations |
"It looks right" is not a validation method. Apply Source Check and Plausibility Scan to every AI output before sharing it.
5Your Action Plan
Techniques only become skills through practice.
| Situation | Use This | Why |
|---|---|---|
| You know what you want and the format is clear | Zero-Shot + COSTARS | Fast, one-shot execution with full coverage |
| You need consistent formatting across items | Few-Shot | Pattern matching eliminates variation |
| You need domain expertise or a specific tone | Role Prompting | Sets vocabulary, depth, priorities |
| A complex decision with multiple factors | Chain-of-Thought | Forces visible, auditable reasoning |
| You're not sure what you need | Interview-Style | AI gathers context through questions |
| You know the goal but can't phrase it | Meta Prompting | AI writes the prompt for you |
By Friday: run the COSTARS prompt from Step 2 today, the Interview-Style prompt from Step 3 tomorrow, then pick the single highest-value opportunity from either output and spend 15 minutes on it with Claude.
Save the outputs and the prompts that generated them. Within a month you'll have a personal library covering your most common work.
Verification
Related Resources
- claude.ai — Run every prompt in this lesson
- Claude documentation — Official product and feature documentation
- Claude support — Help center and account support
Lesson complete
Four techniques, one framework, one exploratory method, and six validation habits — everything you need to use Claude effectively starting today.