Interview Quick Revision

This section is for reading the day before the interview: short answers, comparison tables and a practice checklist.

30.1 Top questions: short answers

30.2 Comparison cheat-sheet

30.3 Code questions that may be asked

# 1. Retry with backoff (the SDK already does this, but here is the manual pattern)
import time, anthropic
def call_with_retry(**kw):
    for attempt in range(5):
        try:
            return client.messages.create(**kw)
        except anthropic.RateLimitError:
            time.sleep(2 ** attempt)              # exponential backoff
        except anthropic.APIStatusError as e:
            if e.status_code < 500: raise         # don't retry 4xx
            time.sleep(2 ** attempt)
    raise RuntimeError("retries exhausted")

# 2. Model router: easy tasks go to the cheaper model
def route(question):
    label = client.messages.create(
        model="claude-haiku-5-5", max_tokens=10,
        system="Write only EASY or HARD.",
        messages=[{"role": "user", "content": question}]).content[0].text
    return "claude-opus-5-5" if "HARD" in label else "claude-haiku-5-5"

# 3. Token count before sending
n = client.messages.count_tokens(model="claude-opus-5-5",
        messages=[{"role": "user", "content": big_text}]).input_tokens

30.4 System design: "Build a chatbot over company docs"

1. Requirements: users, data size, freshness, latency, privacy (PII), cost budget.
2. Ingestion: docs → chunk → contextual embeddings → vector DB + BM25 index.
3. Query: hybrid retrieve → rerank → top chunks + system prompt (cached) → Claude (Sonnet/Opus) with citations.
4. Guardrails: input screening, "answer only from the docs", PII masking, refusal handling.
5. Scale/cost: prompt caching, model routing (Haiku for simple), streaming UX, rate-limit retries.
6. Quality: eval set (100+ real questions), LLM-as-judge, feedback thumbs up/down, monitoring.

30.5 Practice checklist

• Create an API key in the Console and send your first request from both Python and C#
• Build a streaming chat CLI that remembers history
• Write a manual tool loop with one tool (a SQL query or weather)
• Do the same thing with Tool Runner and understand the difference
• Use structured outputs to extract an invoice PDF's data into JSON
• Add prompt caching and verify cache_read_input_tokens
• Run sentiment on 100 reviews with the Batch API
• Build an MCP server with FastMCP and add it in Claude Code with claude mcp add
• Create a CLAUDE.md, a subagent and a PostToolUse hook for your own project
• Write a SKILL.md and test that Claude loads it at the right time
• Build a small RAG bot + an eval set of 20 questions
• Add IChatClient + the tools of an MCP server in a C# console app
• Turn on compaction, run a long conversation and look at the compaction block
• Build a simple file-based backend for the memory tool
• Create a Managed Agent, run a session and view the trace in the Console
• Build a plugin in Claude Code (command + subagent + hook) and install it