kmshihab7878 Building LLM-Powered Applications with Claude
Use this skill for Claude API, Anthropic SDK, and Agent SDK implementation work. Start with the simplest surface that satisfies the task, detect the project language, then load only the relevant language-specific files.
Defaults
Unless the user requests otherwise, use Claude Opus 4.6 with the exact model string claude-opus-4-6, adaptive thinking (thinking: {type: "adaptive"}) for non-trivial work, and streaming for long input, long output, or high max_tokens. Use SDK final-message helpers such as .get_final_message() / .finalMessage() when individual stream events are not needed.
Do not invent model IDs or append date suffixes. If an exact model is not in the cached table in references/operating-reference.md, read shared/models.md or current official docs before using it.
Language Routing
Detect language from project files before loading examples:
| Signals | Read |
|---|---|
Python files, pyproject.toml, requirements.txt |
python/ |
TypeScript/JavaScript files, package.json |
typescript/ |
| Java/Kotlin/Scala files, JVM build files | java/ |
go.mod or Go files |
go/ |
Ruby files or Gemfile |
ruby/ |
| C# project/files | csharp/ |
PHP files or composer.json |
php/ |
| Raw HTTP or unsupported language | curl/ |
If multiple languages are plausible, use the current file/question context or ask one clarifying question.
Surface Selection
Default order:
- Claude API single call for classification, summarization, extraction, Q&A, and simple generation.
- Claude API + tool use for code-orchestrated workflows and custom tools.
- Agent SDK when Claude itself needs built-in file, web, terminal, permission, or MCP capabilities.
- Custom agent loop only when the task is open-ended and the added cost/risk is justified.
Before building an agent, check complexity, value, viability, and cost of error. If any fails, use a simpler surface.
Architecture Rules
- Everything routes through the Messages API unless a supporting endpoint is specifically needed.
- Use SDK tool runners/helpers instead of hand-rolled loops where possible.
- Use
output_config: {format: ...}for structured outputs; do not use deprecatedoutput_format. - Use typed SDK request/response/error types instead of redefining them.
- Do not silently truncate oversized inputs; discuss chunking, summarization, or file APIs.
Reading Guide
Load only what the task needs:
- Basic request/response:
{lang}/claude-api/README.md - Streaming UI:
{lang}/claude-api/README.md+{lang}/claude-api/streaming.md - Tool use / function calling / structured outputs:
{lang}/claude-api/README.md+shared/tool-use-concepts.md+{lang}/claude-api/tool-use.md - Batch processing:
{lang}/claude-api/README.md+{lang}/claude-api/batches.md - Reused file uploads:
{lang}/claude-api/README.md+{lang}/claude-api/files-api.md - Agent SDK:
{lang}/agent-sdk/README.md+{lang}/agent-sdk/patterns.md - HTTP errors:
shared/error-codes.md - Current official docs:
shared/live-sources.md
Read references/operating-reference.md for the full language matrix, decision tree, model table, thinking/effort rules, compaction details, WebFetch guidance, and pitfalls.
Current Critical Pitfalls
- Opus 4.6 / Sonnet 4.6 use adaptive thinking; do not use deprecated
budget_tokensunless an older model is explicitly requested. - Opus 4.6 does not support assistant-prefill control; use structured outputs or prompt instructions.
- Large
max_tokensresponses require streaming to avoid timeouts. - Parse tool-call inputs with JSON parsers, not raw string matching.
- Prefer SDK helpers and SDK-exported types over custom wrappers and duplicate interfaces.
- For reports/documents/visualizations, consider code execution and Files API outputs when appropriate.
When to Fetch Current Docs
Use current official documentation when the user asks for latest/current behavior, cached data appears stale, or the requested feature is not covered locally. Reference URLs are in shared/live-sources.md.