Anthropic's Prompt Engineering Patterns

Anthropic maintains an extensive public prompt engineering guide for Claude (docs.anthropic.com/claude/docs/prompt-engineering). It reflects what they've seen work across millions of production prompts. The patterns transfer to other models, but they are especially potent with Claude.

Pattern 1: Be Clear, Direct, and Detailed

Claude responds best to prompts that read like instructions to a smart new employee. Don't assume context. Spell out the task, the audience, the constraints, and the format. Vague 'help me' prompts produce vague results.

GOOD:
Write a 300-word internal memo announcing a 4-day workweek pilot starting July 1. Audience: 50-person B2B SaaS company, ages 22-55, mostly engineers. Tone: warm but precise. Include: the rationale (work-life balance + productivity research), the logistics (Friday off, same pay, 12-week pilot), and how feedback will be collected. Close with a line inviting questions.

TOO VAGUE:
Write a memo about our new work policy.

Pattern 2: Let Claude Think (Extended Thinking)

For complex reasoning, give Claude explicit room to think before answering. Modern Claude models have an 'extended thinking' mode that surfaces internal reasoning. Even without that feature, structured thinking tags help.

<task>
Decide whether the company should acquire Startup X for $40M.
</task>

<context>
[financials, market data, team bios, etc.]
</context>

Before answering:
1. Inside <analysis> tags, list 5 arguments for and 5 against.
2. Inside <risks> tags, name the 3 highest-impact risks and their mitigations.
3. Inside <recommendation> tags, give your final call with a confidence level (low/med/high).

Use the tags exactly as named.

Pattern 3: Chain Complex Prompts

When a task has multiple independent sub-tasks — research, outline, draft, critique, revise — don't try to do it in one monolithic prompt. Chain them. Each prompt in the chain focuses on one thing and passes its output to the next.

STEP 1 (research):
You are a research analyst. Given the topic '{TOPIC}', list 10 specific claims worth verifying, each with a category tag.

--> output: list of claims

STEP 2 (verification):
Given these claims: {CLAIMS_FROM_STEP_1}, for each one, judge its likely accuracy and mark: supported | contested | unknown. Explain briefly.

--> output: annotated claims

STEP 3 (synthesis):
Using only the 'supported' claims from {ANNOTATED_CLAIMS}, write a 500-word essay with inline citations of the form [claim-id].

--> output: essay

Pattern 4: Give Claude a Role

• 'You are a meticulous copy editor who prefers plain language.'
• 'You are a senior data engineer reviewing a PR.'
• 'You are a contrarian debate partner who argues the strongest opposing case.'
• 'You are a structured-interview AI evaluator scoring a response 1-5 on four dimensions.'

Pattern 5: Use Examples (Multishot)

Anthropic's docs are explicit: 'Multishot prompting is the single most effective technique for improving Claude's performance.' Three well-chosen examples often outperform any amount of verbal instruction.