Claude AI products & certifications portal

Fable 5 is the most capable model — reach for it on the hardest reasoning and long-horizon agentic work. Opus 4.8 is the most capable Opus tier — highly autonomous coding, knowledge work, and memory. Sonnet 5 balances intelligence and cost — near-Opus quality at Sonnet price, the default for most production workloads and agents. Haiku 4.5 is fastest and cheapest — ideal for classification, routing, and sub-agent roles. Strategy: start with Sonnet 5, escalate to Opus 4.8 or Fable 5 where evals show quality gaps, downgrade to Haiku where latency or cost dominates.

Mark stable prefix content — system instructions, large documents, tool definitions — with a cache_control breakpoint. Anthropic stores that prefix server-side and serves later requests with the same prefix at reduced input token cost and lower latency. Use it whenever you have repeated large context: RAG systems, agentic loops with stable tool schemas, and multi-turn chats with fixed system prompts. Cache TTL is 5 minutes by default, refreshed on each hit.

Claude is trained to treat XML as semantic structure. Wrapping content in tags like , , clearly separates context, examples, task, and output requirements. This reduces ambiguity, improves instruction following, and makes outputs easier to parse programmatically when you ask Claude to respond inside specific tags.

Compound several techniques: (1) provide source material in the prompt and instruct Claude to cite it, (2) give Claude explicit permission to say "I don't know," (3) ask it to reason step-by-step before answering, (4) enable extended thinking for complex reasoning, (5) ground claims via retrieval, (6) run judge-model evals and tune. The strongest intervention is usually giving Claude the information it needs rather than relying on parametric memory.

The system parameter sets persistent behavior, role, constraints, and tone — outside the messages array. The messages array holds alternating user/assistant turns. Best practice: stable instructions, persona, output format, and policies go in system; actual tasks, data, and conversation history go in messages. System content also caches well.

CLAUDE.md is a project memory file Claude Code reads automatically at session start. Put your project's architecture, coding conventions, key directories, important commands, and team rules inside it. Every new conversation begins with this context loaded — eliminating repetitive setup. A user-level file lives at ~/.claude/CLAUDE.md; project-level files override user-level.

Use /compact to summarize prior context instead of /clear when you want to keep a thread. Write precise prompts targeting specific files. Move recurring instructions into CLAUDE.md. Limit simultaneously enabled MCP servers — too many tool definitions can shrink usable context from 200K to under 100K tokens. Use sub-agents for isolated explorations that would otherwise pollute the main session.

Hooks are shell commands triggered by Claude Code lifecycle events — before/after a tool call, before a commit, on session start. Use them to enforce linting, run tests, log actions for audits, block dangerous commands, or notify CI systems. They're defined in settings.json and let you wrap Claude's autonomy in deterministic guardrails.

MCP is an open standard for connecting AI agents to external tools and data sources. An MCP server exposes tools, resources, and prompts; a client (Claude Code, the API's MCP connector, or the Claude.ai interface) discovers and calls them. Transports include stdio (local processes), SSE, and HTTP. It eliminates one-off integrations — any MCP-compliant server works with any MCP-compliant client.

Treat MCP servers like any privileged subprocess: install only from trusted sources, use scoped credentials with minimum required permissions, store secrets in env vars or a secrets manager (not plaintext configs), enable structured logging for all tool calls, restrict network egress, and run processes in isolated containers for high-risk integrations. Pin server versions to avoid supply-chain surprises.

Start with explicit success criteria. Build the tool surface (typed JSON schemas, descriptive names, narrow scope per tool). Pick a model — usually Sonnet for the orchestrator, Haiku for cheap sub-tasks. Implement the loop: prompt → model → tool call → execute → feed result back → repeat until done or max iterations. Add guardrails: step cap, cost ceiling, timeout, observability. Write evals before you scale. Parallelize with sub-agents.

Sub-agents are isolated Claude instances with their own context window, spawned via the Task tool to handle a sub-task, returning only the useful result. Use them for codebase exploration, code review, parallel research, or any work that would bloat the main session. Define them in .claude/agents/*.md with their own system prompt, tool restrictions, and permission mode.

Implement exponential backoff with jitter on 429 and 5xx responses. Respect the retry-after header. Use the Batch API for non-real-time workloads — ~50% cheaper and on a separate quota. Spread load across model tiers. Cache stable prefixes to reduce token throughput. Monitor API usage per key to catch runaway loops early.

Treat all untrusted input (user messages, web content, file contents, tool outputs) as data, not instructions. Wrap it in XML tags and explicitly tell Claude not to treat content inside as instructions. Constrain tools — never give destructive actions without human-in-the-loop confirmation. Use allow-lists for URLs and commands. Run evals with adversarial test cases. For high-stakes apps, add a classifier to flag suspicious inputs before they reach Claude.

Extended thinking gives Claude a dedicated internal reasoning budget before producing its final answer. Set a thinking.budget_tokens parameter to control how long Claude reasons. Use it for tasks that benefit from deeper deliberation: hard math, complex logic, multi-step planning, or any situation where chain-of-thought alone isn't sufficient. Not recommended for simple tasks — it adds latency and cost.