XML tags, prompt chaining, meta-prompting, output constraints, and Anthropic's official pattern library. These techniques separate professional Claude users from power users — the difference between good outputs and consistently excellent outputs.
Anthropic's official documentation recommends using XML-style tags to clearly separate different components of complex prompts. This dramatically improves how Claude parses and responds to long, multi-component prompts.
Claude was trained on internet text that included HTML/XML, so it understands that content within tags is semantically distinguishable. Tags create a clear visual and semantic boundary between different types of information in a prompt — context vs. instructions vs. data vs. expected output format. This reduces Claude's risk of "confusing" which part of a long prompt is the task vs. the background.
<context> I am a college junior majoring in Economics conducting research on behavioral economics for my thesis. My advisor specializes in decision theory. I have a strong statistics background. </context> <task> Explain the difference between prospect theory and expected utility theory. Focus on how each predicts behavior differently in loss-aversion scenarios. </task> <requirements> - Use one concrete numerical example for each theory - Explain the mathematical difference in how probabilities are "weighted" - Mention at least two empirical studies that support prospect theory - Keep the explanation under 500 words </requirements> <output_format> Use headers: "Expected Utility Theory" | "Prospect Theory" | "Key Differences" | "Empirical Evidence" </output_format>
<document> [Paste the document, paper, or text to be analyzed here] </document> <analysis_task> Based ONLY on the document above (do not use outside knowledge): 1. What is the central argument? 2. What evidence does the author provide? 3. What assumptions underlie the argument? 4. What is NOT addressed that should be? </analysis_task> <grounding_instruction> If any of your points cannot be directly supported by a quote or specific reference to the document, say so explicitly. </grounding_instruction>
| Tag | Use For | Example |
|---|---|---|
<context> | Background information | Your background, the situation, project details |
<task> | The specific request | What you want Claude to do |
<document> | Content to analyze | Papers, emails, code, articles |
<requirements> | Constraints and rules | Length limits, tone, format rules |
<example> | One-shot/few-shot examples | Input→output demonstrations |
<output_format> | Response structure | Headers, sections, specific format |
<thinking> | Invite CoT reasoning | "Use these tags to think through..." |
<data> | Input data to process | CSV data, user records, test cases |
Prompt chaining breaks complex tasks into a sequence of simpler prompts where the output of each becomes the input of the next. This is one of the most powerful techniques for production-quality work.
CHAIN OVERVIEW: Input (research topic) → [Step 1] → [Step 2] → [Step 3] → [Step 4] → Final Paper Step 1 — RESEARCH OUTLINE PROMPT: "Given my topic [X], create a detailed research outline with: thesis, 5 main arguments, key evidence needed for each, and 3 potential counterarguments." → OUTPUT: Structured outline Step 2 — EVIDENCE GATHERING PROMPT: [Paste Step 1 output] "For each argument in this outline, suggest 2-3 types of evidence I should look for and keywords for Google Scholar." → OUTPUT: Research directions Step 3 — SECTION DRAFTING PROMPT: [Paste outline + evidence directions] "Write the first draft of Section 2 ([title]) using this outline. Target: 600 words, academic tone, leave [SOURCE NEEDED] placeholders where I need to add real citations." → OUTPUT: Draft section Step 4 — REVISION PROMPT: [Paste draft section] "As a strict academic editor, critique this section and rewrite it to address the 3 most important weaknesses you identified." → OUTPUT: Polished section
Meta-prompting is asking Claude to write, evaluate, or improve prompts — for itself or for other tasks. This is a fast way to generate better prompts when you're not sure how to phrase something.
Output engineering means precisely specifying the structure, format, and content of Claude's response to eliminate editing and post-processing work.
For developer use cases, you can instruct Claude to produce structured JSON output that feeds directly into code:
In the Claude API, you can "prefill" the beginning of Claude's response — starting it with specific text that forces the output to begin a certain way. This is powerful for controlling format.
response = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
messages=[
{"role": "user", "content": "Analyze this business case: [case]"},
# Prefill Claude's response to force a specific opening
{"role": "assistant", "content": "```json\n{"}
]
)
# Claude will continue from the prefill, producing pure JSON output
# This eliminates the "here is the JSON:" preamble Claude often adds
Grounding techniques constrain Claude to only use information you've provided — preventing hallucination for high-stakes tasks:
Sometimes you need Claude to absolutely, definitely follow a rule. Use clear constraint language:
"Under no circumstances provide [X]. This is a hard requirement that overrides everything else."
"Every response MUST include [X]. If you cannot include it, explain why rather than omitting it silently."
"Your entire response must be under 100 words. If you cannot answer in 100 words, give the most important 100 words and note what was cut."
"Limit your analysis strictly to [domain]. If asked about unrelated topics, redirect: 'That's outside my focus for this conversation.'"
For high-stakes work (thesis chapters, important code, major decisions), use a structured multi-pass approach where each pass focuses on a different dimension:
Anthropic maintains an official Prompt Library with tested, production-quality prompts. Key patterns from their library that students should know:
Reliable JSON extraction from unstructured text — useful for processing survey responses, classifying emails, extracting entities from documents.
Divergent thinking prompts that prevent Claude from anchoring on the first idea — generates genuinely diverse options rather than variations of one idea.
Patterns that explicitly ask Claude to separate high-confidence claims from uncertain ones — critical for any factual output.
Breaking complex risky tasks into verifiable steps — each step is simpler, safer, and can be checked before proceeding.
Advanced prompting isn't about knowing the most techniques — it's about having a reliable debugging process when outputs aren't right. The expert mental model: