# Structured Thinking XML tags, scratchpads, and extended thinking for deeper model reasoning. ## Why Structure Matters Unstructured prompts lead to unstructured thinking. Giving the model explicit sections for different types of content — context, instructions, scratchpad, output — produces more reliable, higher-quality results. ## XML Tag Delimiters (Claude) Claude is trained to respect XML tags as semantic boundaries: ```xml You are an expert code reviewer. This is a Node.js Express app using Prisma with PostgreSQL. Coding conventions: TypeScript strict mode, named exports, zod validation. {{CODE}} Review the code above. Focus on security and correctness. Output findings as a numbered list ordered by severity. ``` Benefits: cleaner separation of concerns, easier to update individual sections, reduces prompt injection risk. ## Scratchpad Pattern Give the model space to think before producing output: ```xml Is this argument logically valid? {{ARGUMENT}} First, work through your analysis in a . Identify premises, logical structure, and potential flaws. Then give your final verdict in . ``` Model output: ```xml Premise 1: All A are B Premise 2: X is A Therefore X is B This is valid modus ponens... The argument is logically valid. If the premises are true, the conclusion follows necessarily. ``` ## Extended Thinking (Claude) For Claude models with extended thinking enabled, you can request deeper reasoning: ``` Think deeply about the architecture tradeoffs before recommending an approach. Consider at least 3 alternatives and their long-term implications. ``` In the API, enable via: ```json { "thinking": { "type": "enabled", "budget_tokens": 10000 } } ``` Thinking tokens are internal — they don't appear in the output but improve answer quality significantly for hard problems. ## Structured Output with Schema Force a specific output format: ```xml Analyse the code and return a JSON object matching this schema exactly: { "summary": "string — 1 sentence", "severity": "critical" | "high" | "medium" | "low", "issues": [ { "type": "string", "location": "file:line", "description": "string", "fix": "string" } ] } Return only the JSON. No explanation outside the JSON. ``` ## Combining Techniques ```xml You are a principal engineer reviewing architecture proposals. {{PROJECT_CONTEXT}} {{ARCHITECTURE_PROPOSAL}} In , think through the proposal's strengths, weaknesses, scaling implications, and operational complexity. Then in , give a clear recommendation: approve / approve-with-changes / reject. In , list specific required changes if applicable. ``` ## Tips - Use consistent tag names across your prompts — pick a convention and stick to it - Don't over-structure simple prompts — XML tags are overhead for one-liners - Claude handles XML naturally; GPT models may need JSON-style structure instead - Always test whether the model respects your tags with adversarial inputs

Prompts