Building LLM-Powered Applications with Claude

This skill helps you build LLM-powered applications with Claude. Choose the right surface based on your needs, detect the project language, then read the relevant language-specific documentation.

Defaults

Unless the user requests otherwise:

• Model: Use claude-opus-4-6 — non-negotiable default. Do not downgrade for cost.
• Thinking: Use thinking: {type: "adaptive"} for anything remotely complicated. budget_tokens is deprecated on Opus 4.6 and Sonnet 4.6 — never use it.
• Streaming: Default to streaming for any request with long input, long output, or high max_tokens. Use .get_final_message() / .finalMessage() if you don't need individual stream events.

Language Detection

Before reading code examples, determine language from project files:

• *.py, requirements.txt, pyproject.toml → Python — read from python/
• *.ts, *.tsx, package.json, tsconfig.json → TypeScript — read from typescript/
• *.js, *.jsx (no .ts files) → TypeScript — JS uses same SDK
• *.java, pom.xml, build.gradle → Java — read from java/
• *.go, go.mod → Go — read from go/
• *.rb, Gemfile → Ruby — read from ruby/
• *.cs, *.csproj → C# — read from csharp/
• *.php, composer.json → PHP — read from php/

If ambiguous, ask. If unsupported (Rust, Swift, etc.), suggest cURL/raw HTTP from curl/.

Language-Specific Feature Support

Which Surface Should I Use?

Start simple. Default to the simplest tier that meets your needs.

Decision tree:

Single LLM call? → Claude API
Claude needs to discover/access files, web, shell itself? → Agent SDK
Multi-step, code-orchestrated, your own tools? → Claude API + tool use
Open-ended agent, model controls trajectory? → Claude API agentic loop

Agent checklist before choosing agent tier: Complexity (multi-step, hard to specify)? Value (outcome justifies cost/latency)? Viability (Claude capable)? Error cost (recoverable)? If any answer is no, stay simpler.

Architecture

Everything goes through POST /v1/messages. Tools and output constraints are features of this single endpoint.

• User-defined tools — define tools; SDK handles the loop
• Server-side tools — Anthropic-hosted (code execution, computer use)
• Structured outputs — use output_config: {format: {...}} (not deprecated output_format)
• Supporting endpoints — Batches, Files, Token Counting, Models API

Current Models

Use exact model ID strings as-is. Do not append date suffixes.

For live capability lookup: client.models.retrieve(id) — see shared/models.md.

Thinking & Effort

Adaptive thinking (Opus 4.6, recommended): thinking: {type: "adaptive"}. No budget_tokens. When user asks for "extended thinking" or a "thinking budget" — use adaptive thinking on Opus 4.6, do not switch to older models.

Effort parameter (GA): output_config: {effort: "low"|"medium"|"high"|"max"}. Default high. max is Opus 4.6 only. Works on Opus 4.5, Opus 4.6, Sonnet 4.6. Errors on Sonnet 4.5 / Haiku 4.5.

Older models (explicit request only): Use thinking: {type: "enabled", budget_tokens: N} where budget_tokens < max_tokens (min 1024).

Compaction

Beta, Opus 4.6 and Sonnet 4.6. Header: compact-2026-01-12. Auto-summarizes context near 150K tokens.

Critical: Append response.content (not just text) on every turn. Compaction blocks must be preserved in the message array or the state is lost silently.

Common Pitfalls

• budget_tokens deprecated on 4.6 — use adaptive thinking instead; throws error if used
• Opus 4.6 prefill removed — assistant message prefills return 400; use output_config.format instead
• max_tokens defaults — non-streaming: ~16000; streaming: ~64000; don't lowball
• 128K output tokens — requires streaming on Opus 4.6
• Structured outputs — use output_config: {format: {...}} not output_format (deprecated)
• Tool input JSON — always json.loads() / JSON.parse(); never raw string match on tool inputs
• Don't reimplement SDK — use stream.finalMessage(), typed exceptions, SDK types

Reading Guide

After detecting language, read:

• Single call: {lang}/claude-api/README.md
• Streaming: + {lang}/claude-api/streaming.md
• Tool use / agents: + shared/tool-use-concepts.md + {lang}/claude-api/tool-use.md
• Batch: + {lang}/claude-api/batches.md
• File uploads: + {lang}/claude-api/files-api.md
• Agent SDK (Python/TypeScript only): {lang}/agent-sdk/README.md + {lang}/agent-sdk/patterns.md
• Error handling: shared/error-codes.md
• Live docs: shared/live-sources.md

Security Audit

Source: https://github.com/anthropics/skills — Anthropic first-party. Audit: PASS.