Efficient Prompt Engineering Guide

Core mindset: Prompt is code — not a magic word, not a chatbot message. It must be structured, version-controlled, and iteratively tested.

Framework 1 — C-T-R Formula

Every effective prompt needs 3 parts:

• C — Context: Who you are, background, project info
• T — Task: The specific action you want done
• R — Rules: Tone, format, constraints, exclusions

Talk to Claude like briefing a smart new colleague — natural, concise, with enough context. No "keyword magic" needed.

Framework 2 — 3 Layers of Description

Most people only specify Product. Good prompters add Process. Fluent prompters nail all three — especially Performance, which is most often skipped.

Weak prompt (Product only):

Write an email to a customer complaining about a bug.

Strong prompt (all 3 layers):

[Product] Write a professional email (≤150 words, markdown not needed) 
to a customer who reported Bug #4521 in our payment module.

[Process] Step 1: Acknowledge the issue with empathy. Step 2: State the 
fix ETA (48 hours). Step 3: Offer a 10% discount coupon as goodwill.

[Performance] Tone: warm but professional. Do NOT over-apologize. 
Ask me before sending if the ETA needs to change.

The 6 Core Techniques

You don't need all 6 in every prompt. Pick based on the task.

1. Give Context (W5H)

AI knows a lot but knows nothing about your specific context. You must feed it.

Structure your context using W5H:

• What — the specific task
• Why — the purpose
• Who — audience / stakeholder
• When — deadline / timing
• Where — the channel or platform
• How — how the output will be used

❌  "Write an announcement for the team"

✅  "Write a Slack announcement for our 15-person engineering team 
    (non-technical managers included). Purpose: inform them that 
    the production deploy is delayed by 2 hours due to a hotfix. 
    Tone: calm, transparent. Max 3 sentences."

When to use: Task has stakeholders, domain-specific context, or history.

2. Show Examples (Few-Shot Prompting)

One example = 10× better than an abstract description. Show, don't tell.

• Provide 2–5 input → output examples
• AI pattern-matches and produces output matching your examples
• Corner case examples are the most powerful — they teach nuance

When to use: Style-heavy tasks, pattern repetition (taglines, headlines), when you have good past outputs.

3. Specify Constraints

AI tends to over-engineer. Constraints make it focus.

Categories to constrain:

• Length: "exactly 50 words", "≤3 bullets"
• Format: "markdown table", "JSON only", "plain prose"
• Must-include: "MUST contain keyword X"
• Must-exclude: "MUST NOT mention competitor Y"
• Tone boundary: "playful but professional"

Positive constraints are easier for AI to follow than negative ones.
Start simple — only add constraints when you see the output going wrong.

4. Break Complex Tasks into Steps

LLMs struggle with implicit multi-step reasoning. Make it explicit.

❌  "Analyze competitors and recommend a strategy"

✅  "Analyze competitors and recommend a strategy using these steps:
    1. List 5 direct competitors for Product Y
    2. Profile each: offering, pricing, USP, customer reviews
    3. Build a comparison matrix (5 competitors × 8 dimensions)
    4. Identify gaps (where all competitors are weak → opportunity)
    5. Identify saturation (where all are strong → table stakes)
    6. Recommend 3 strategies based on gap analysis (rationale + risk each)
    7. Pick the top 1 with full justification"

Benefits: Easier to verify each step, reduces reasoning drift, transparent thinking.

5. Ask AI to Think First

Force the AI to output reasoning before the final answer — this drastically reduces hallucinations.

4 ways to trigger thinking:

1. "Before answering, think step by step. Put your reasoning in <thinking> tags, then give your final answer."
2. "Before writing the full essay, share your outline for me to review."
3. "Before solving, list 5 assumptions or questions you need to clarify."
4. Enable Extended Thinking mode in Claude (for complex tasks).

When to use: Reasoning-heavy tasks, high-stakes output, ambiguous data.

6. Define Role & Tone

The same question asked to different personas gives vastly different answers. Calibrate the AI with a Role + Audience + Tone combo.

Level 1 — Role:    "You are a senior financial analyst with 15 years 
                    covering the tech sector."
Level 2 — Tone:    "Tone: professional, analytical, no fluff. 
                    Use data whenever possible."
Level 3 — Audience:"Audience: non-technical CEO with a 5-minute 
                    attention span. Avoid jargon, use analogies."

Advanced Strategies

Clear & Direct: The First Line Rules Everything

The first line is the most important. Claude orients from line 1.

❌  "Could you maybe write something about photosynthesis for students?"
✅  "Generate a 10-question quiz on photosynthesis for Grade 9. 
    Mix: 5 multiple-choice, 3 short-answer, 2 essay. Include an 
    answer key with explanations."

Structure with XML Tags

For prompts 20+ lines or mixing instructions with data, use XML tags to help Claude parse boundaries.

<task>
Summarize the following customer reviews and classify sentiment.
</task>

<brand_guidelines>
- Tone: empathetic and solutions-focused
- Never mention competitor names
</brand_guidelines>

<reviews>
Review 1: "The product is great but shipping took 3 weeks..."
Review 2: "Customer support was amazing when I had an issue..."
</reviews>

<output_format>
Return a JSON with: summary (string), sentiment (positive/neutral/negative), 
top_issue (string)
</output_format>

Rules: Max 2 levels of nesting. Use descriptive tag names (<sales_data> not <data>). Only use XML for complex prompts — overkill for short ones.

Context Engineering: Place Information Strategically

LLMs have a "lost in the middle" effect — attention is strongest at the beginning and end, weakest in the middle.

Stanford 2023 research: Accuracy drops >30% for facts placed in the middle of long prompts.

Rules:

1. Place critical instructions at the top (system prompt) and repeat them at the end
2. Curate ruthlessly — 5 relevant pages > 50 pages of "all the info I have"
3. Every extra context you add pushes other context into the attention dead zone

IPO Framework: Intent → Purpose → Output

Avoids the "letter over spirit" failure — where AI follows your literal words but misses your actual goal.

❌  "Make this email shorter."  
    (AI deletes the call-to-action — technically shorter, useless)

✅  "Make this email shorter. 
    Goal: keep executive attention while preserving the final 
    recommendation and the Friday deadline."

Rule: Any time you use a comparative ("shorter", "better", "more X") — always add what goal it serves.

Meta-Prompting: Use AI to Improve Your Own Prompt

The most powerful technique that 90% of people ignore.

Paste your draft prompt and ask:

"Identify what is missing or ambiguous in this prompt. 
Suggest specific improvements."

Why it works:

• AI sees ambiguity you miss (you're too close to the context)
• AI knows what detail models typically need
• AI can compare your prompt against successful patterns in its training data

Eval-Driven Development: The 5-Step Loop

True prompt engineering requires measurement, not just gut feel.

Set Goal → Write v1 → Evaluate → Apply ONE Technique → Re-evaluate → Repeat

Real example score progression:

Anti-Patterns to Avoid

Quick-Reference Checklist

Before sending a complex prompt, check:

• First word is a strong action verb?
• Context covers W5H?
• Constraints specified (length, format, must-include/exclude)?
• Complex task broken into numbered steps?
• Examples provided for style/pattern tasks?
• Role + Audience + Tone defined?
• Critical info at the beginning AND end (not buried in the middle)?
• Did you ask AI to think first for reasoning-heavy tasks?
• Did you run meta-prompt to catch blind spots?

Model Selection Guide

Source: Claude Course 2026 — NotebookLM notebook
Notebook: Open in NotebookLM