Anthropic Tool-Use API

Overview (confidence: high)

Anthropic’s tool-use API set the standard that most coding tools follow. It uses a narrow XML envelope (3-5 top-level tags) with separate chat and tool-use prompts. It is the originator and dominant user of the xml-envelope-format in the corpus.

Envelope Format (confidence: high)

Narrow XML envelope with minimal tag hierarchy. Tool descriptions use <tool> blocks with <parameters> sub-blocks, establishing the format that tools like cline-system-prompt and Comet adopted. The narrow structure optimizes for LLM parsing clarity.

Constitutional AI Layer (confidence: high)

Anthropic implements the only Level 4 defense in the corpus — a constitutional AI layer that evaluates outputs before delivery. This prevents prompt extraction via self-evaluation. It is computationally expensive but provides the strongest protection against extraction attempts. See anti-prompt-injection for the four-level defense taxonomy.

Patterns Used (confidence: high)

• Implicit chain-of-thought: Relies on the model’s native reasoning ability rather than explicit CoT instructions
• Discussion-before-code default: Asks clarifying questions before implementing
• Multi-identity persona: “You are a helpful assistant AND a code editor”
• Crisis detection: Explicit detection of self-harm, self-injury, or suicidal content with crisis resource provision

Communication Style (confidence: high)

Follows explain-before-do pattern. The tone is pedagogical and thorough. Designed for complex reasoning tasks where explanation builds trust.

Guardrails (confidence: high)

Direct refusal pattern for harmful requests. The constitutional AI layer provides self-evaluation before delivery. Crisis detection is one of only two tools in the corpus (with Claude Code) that explicitly handles safety-critical user states.

Implications (confidence: high)

Anthropic’s tool-use API set the standard that most tools follow. Its narrow XML envelope optimizes for LLM parsing clarity. The constitutional AI layer is the only Level 4 defense in the corpus — computationally expensive but strongest protection.

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]