Building with the Claude API

API

Auth, the messages array, streaming, token limits, structured output, and the errors you will hit in week one. Everything else you can look up later.

System Prompt

The system prompt is the highest-leverage thing you control. What breaks when you get it wrong, and the patterns that hold up in production.

Streaming

When to stream, SSE to the browser, error handling mid-stream, and the UX patterns that stop your interface from feeling broken.

RAG

Chunking strategy, embedding choice, hybrid retrieval, reranking, and how to measure whether the pipeline is actually working.

Tool Use

How Claude calls functions you define. The tool definition format, parsing tool-use responses, handling multi-step tool calls, and the failure modes.

Evals

Test case structure, deterministic vs. LLM-as-judge assertions, running evals in CI, and the number to track over time.

Prompt Caching

One parameter change that cuts cost 80%+ on workloads with repeated context. Multi-block caching, conversation history caching, verifying hits in usage data, and the gotchas that eat your savings.

Cost Optimization

Measurement first. Then: model routing by task complexity, batch API for async workloads, output length control, and context window management. Most apps find 40-60% savings in the first audit.

Tool Use

Defining tools that Claude calls correctly, handling multi-turn and parallel tool calls, returning errors cleanly, and the failure modes that will bite you in production.

Multi-agent System

Orchestrator + subagent patterns, parallelism with ThreadPoolExecutor, state management across agents, checkpoints, and the patterns that look good until production.

Multi-agent Evaluation

Golden datasets, LLM-as-judge pairwise scoring, ablation testing to find the agent adding only latency, and the cost/quality equation that determines if orchestration is worth it.

Persistent Memory

User facts, session summaries, entity notes. How to extract, store, and inject memory without token bloat — and the difference between what is worth remembering and what is not.

Production Deployment

Secrets management, rate limit handling, cost controls, observability, and the pre-launch checklist. What separates a localhost demo from a production app.

Production Monitoring

The logging wrapper, the five metrics that matter, what Anthropic rate limiting looks like in practice, cost alerts, and the four dashboard questions you should be able to answer by month one.

Error Handling

API error taxonomy, output format resilience, hallucination detection, context overflow guards, and user-facing error patterns. The full error surface of a Claude app.

Security

The four security issues that show up most often in Claude apps: API key exposure, prompt injection, over-permissioned tools, and untrusted system prompt content. With a pre-deploy checklist.

Authentication

NextAuth setup, protecting API routes with middleware, passing user identity into Claude calls, and the session provider pattern for App Router.

Conversation Persistence

Schema design, loading history server-side, pruning context intelligently, and the URL-based conversation pattern that makes chats resumable and debuggable.

Rate Limiting

Token budgeting, per-user application-layer rate limiting, request queuing for burst scenarios, and graceful degradation patterns.

Full-Stack Chatbot

Everything in this path combined: Next.js App Router, Claude API with streaming, NextAuth authentication, Supabase message persistence, and deployment to Vercel. Zero to shipped.