Anthropic★ 4.3

Claude API

Anthropic's official Skill for building with the Claude API/SDK. Patterns for caching, streaming, tool use, model migration.

View repo Homepage

A+

Safety & Trust

Delv Safety Grade: A+

Score 94/100 · assessed 2026-04-18

Maintainer100

Permissions95

Supply chain90

Transparency90

Incidents100

This is Anthropic's first-party Skill for teaching Claude correct API usage patterns. It's a reference bundle, not executable code, so the attack surface is minimal. The Skill provides working examples for prompt caching, streaming, tool use, and model migration, reducing hallucinated SDK patterns. Maintained directly by Anthropic in their official skills repository with clear documentation and versioning. The primary risk is that Claude might over-rely on these patterns when fresher approaches exist, but that's a design tradeoff rather than a security concern. Supply chain is clean: official GitHub repository, no external dependencies, distributed as documentation rather than executable packages. Transparency is excellent with open source code and clear examples. No known incidents. This is as safe as agent skills get.

Green flags

Official Anthropic first-party Skill, highest trust anchor
Read-only reference material, no code execution or API calls
Open source with clear examples in public GitHub repository
Addresses real problem: SDK hallucination and incorrect API patterns
No dependencies, no install process, pure documentation

Red flags

Documentation-based Skill could become stale as API evolves
No automated testing visible for example code accuracy

Permissions requested

Outbound networkExternal LLM call

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

This Skill teaches Claude how to write correct, production-ready code using Anthropic's official SDK. It covers the fiddly bits that trip people up: structuring prompt-cached requests so they actually cache, setting up tool schemas that Claude can call reliably, handling streaming responses without dropped events, and migrating between Sonnet versions without breaking existing integrations. The Skill bundles working examples and patterns from Anthropic's own documentation, so when you ask Claude to add caching to an existing app or debug a streaming integration, it references real code instead of hallucinating plausible-looking patterns that fail at runtime. Particularly useful for anyone building apps that need to manage API costs through caching or expose tools to Claude in a structured way. It won't help with raw HTTP calls or third-party wrappers, but if you're using the official Python or TypeScript SDK, it's a solid reference that keeps Claude from inventing its own creative interpretations of the API.

Review

I've watched Claude hallucinate SDK patterns enough times to appreciate what this Skill does. It's a reference bundle that teaches Claude how to write correct Anthropic API calls, handle prompt caching, stream responses properly, and migrate between model versions without breaking things. Before loading this, I'd ask Claude to add tool use to an existing app and get back code that looked plausible but failed at runtime because it misunderstood the Messages API structure or forgot to set up the tool schema correctly. With the Skill loaded, Claude references actual working patterns from Anthropic's own examples. The difference shows up most clearly in caching scenarios. Prompt caching can cut costs by 90% if you structure requests correctly, but the rules are fiddly: cache breakpoints need to sit at specific boundaries, you have to mark blocks explicitly, and the minimum cacheable size matters. Without this Skill, Claude often generates code that looks right but doesn't actually cache anything. With it, I get working cache-aware code that matches the official docs. The migration guidance is surprisingly useful too. Moving from Sonnet 4.6 to 4.7 isn't just a model name swap, there are behaviour changes and new parameters. The Skill walks Claude through what actually changed and how to update code safely. Streaming is another win. Claude now generates proper async handlers that don't swallow errors or leave connections hanging. The rough edge is that this only helps if you're building with Anthropic's SDK specifically. If you're hitting the API directly via HTTP or using a third-party wrapper, the patterns don't transfer cleanly. And it's overkill for trivial single-shot API calls. But for anything involving caching, tools, or streaming, I load this Skill first and save myself a debugging session.

Verdict

Load this if you're building anything non-trivial with the Claude API: caching, streaming, tool use, or model migrations. Skip it if you're just making basic one-off completions.

Good at

Teaches Claude the correct prompt caching structure, which can cut API costs by 90%
Provides working streaming patterns that handle errors and connection cleanup properly
Includes migration guidance for moving between Sonnet versions without breaking changes
References official Anthropic examples instead of hallucinated patterns
Covers tool use setup comprehensively, including schema structure and error handling

Watch out

Only useful if you're using Anthropic's official SDK, not raw HTTP or third-party wrappers
Overkill for simple single-shot API calls without caching or streaming
Doesn't cover non-Claude models or multi-provider setups
Limited to API integration patterns, won't help with prompt engineering or application architecture

Use cases

Writing prompt-cached SDK calls correctly
Migrating from Sonnet 4.6 to 4.7
Adding tool use to an existing app
Debugging a streaming integration