Skip to content
Hari KharelStudy

Anthropic

Practical, multi-level coding with no AI: you build one small system and it grows over 3-4 levels, mostly in Python. Speed decides it, so keep every level working before you add the next. On top of that come system design (often LLM-serving shaped), a project deep-dive, and a values round. Recruiters name the values round as the one most candidates fail.

The loop

Section titled “The loop”
Stage Format Evidence
Recruiter call 15-30 min: background, level, “why Anthropic” reported
Coding screen 60-90 min. Either a CodeSignal project with 4 levels that extend one codebase (camera on), or a live screen in CodeSignal or Google Colab over Google Meet reported
Hiring manager 45-60 min resume walkthrough, sometimes with code review. Some loops fold it into the onsite reported
Virtual onsite About 4-5 one-hour rounds: 1-2 coding, system design (shared Google Doc), project deep-dive, values/culture, behavioral reported
References, then team match References are taken seriously; rejections at this stage are reported reported
  • Tools. Google Meet, with “Colab and CodeSignal” for technical roles. “You can look things up, just be comfortable with basic syntax and standard libraries.” official
  • AI. Live interviews: “This is all you, no AI assistance unless we indicate otherwise.” Take-homes: no Claude unless told otherwise. Using Claude to prepare is fine. official
  • Language. Python dominates reports. One 2026 report describes a TypeScript-only round where switching language was refused. thin Ask whether the language is fixed per round.
  • Timeline 3-6 weeks. No feedback on rejections. official

What the coding rounds test

Section titled “What the coding rounds test”

Style. Not LeetCode puzzles. You get a realistic spec, implement Level 1, and then each level adds requirements on top of the same code. On CodeSignal, a level must pass hidden tests before the next one unlocks. Running out of time is the most common complaint in reports. reported

Problem family candidates report How well attested Study first
In-memory database / KV store over levels (queries, then TTL, history, or backup and restore) Most consistent family; Glassdoor and several prep sites Incremental requirements, Class design, TTL and expiry, Snapshots
Banking / transactions with levels reported Money arithmetic, Scheduling
Web crawler: serial, then concurrent without visiting a page twice reported Concurrency “recurs across rounds” Crawling, Concurrency, BFS
Profiler / stack-sample processing (samples to a call trace, longest-running function) reported Stack, Simulation
Duplicate-file detection over a directory tree reported Hash maps, File paths, Memory efficiency
Image-transform pipeline driven by a spec thin ranked #1 on one prep site’s list API design, Input validation
LRU cache with extra constraints thin Caching, Linked lists

The skill underneath them all: structuring a class so Level 3 doesn’t force a rewrite of Level 1, plus Python stdlib fluency (dict, collections, heapq, bisect, threading/queue or asyncio).

System design

Section titled “System design”

Either design from a blank prompt or critique an existing design doc. reported Reported topics:

  • LLM inference serving: batching, queues, GPU utilisation
  • chat or messaging systems
  • distributing large model files; batch model runs
  • a prompt playground; analytics for a chat app
  • a request-routing or scheduling layer

One prep site notes that “most AI-specific prompts reduce to standard distributed systems problems.” Expect requirements, APIs, data model, scale, failure and recovery, overload protection, and trade-offs. Read Rate limiting, Scheduling, Resource allocation and Distributed thinking.

Values and project rounds

Section titled “Values and project rounds”
  • Values / culture. Mission, AI safety, ethical dilemmas with no clean answer, and a balanced critique of Anthropic itself. Follow-ups get personal (“how did that feel?”). Candour beats agreement; rehearsed or uncritically enthusiastic answers are a named failure mode. reported Anthropic calls itself “high-trust, low-ego” and lists values such as “put the mission first”. official Read Core Views on AI Safety and form your own view, including where you disagree.
  • Project deep-dive. One large project you owned. Separate your decisions from the team’s, and cover trade-offs and what you’d change. Resume claims that don’t survive deep questions sink candidates. reported

How you’re judged, and why people fail

Section titled “How you’re judged, and why people fail”

Anthropic publishes no rubric: “We care about what you can do, not where you learned to do it.” official Candidates consistently report being judged on:

  • working code that passes at each level
  • adapting cleanly to new requirements
  • naming edge cases before you’re prompted
  • explaining your decisions

Failure reasons that recur:

  1. The values round: flat, rehearsed or gushing answers; no real “why Anthropic”. (Recruiters quoted by one source say it’s where most candidates fail.)
  2. Time: not reaching the later levels, usually after over-engineering Level 1 or rewriting at Level 3.
  3. Weak concurrency: threads, locks, queues, avoiding duplicate work.
  4. A perfect screen isn’t a pass: some candidates with a perfect coding score were still rejected on other rounds.

Your plan

Section titled “Your plan”

Use the Sandbox’s Anthropic filter for practice. Of its 38 problems, the ones below match the reported families most closely. Work in a browser editor with autocomplete off, a visible timer, and no AI.

2 weeks 4 weeks
Week 1 Multi-level builds: Versioned Record Store, Bank, Task Manager. Then Web Crawler. Write your values notes and one project story Week 1 of the 2-week plan, plus Cloud File Storage and Transactional Key-Value Store
Week 2 Profiler from Sampled Stacks, Duplicate Files, Image Pipeline, then one timed mock with Mini Redis (stop at 90 min). Two design sessions: inference serving, chat. Rehearse values answers out loud Concurrency: Web Crawler again from scratch, Blocking Queue Simulator. Profiler from Sampled Stacks, Duplicate Files
Week 3 Image Pipeline, LRU Cache, Inference Batcher. Three design sessions: inference serving, chat, model-file distribution
Week 4 Two full timed mocks (Mini Redis, Access Request System). Values and deep-dive mocks with a friend. Rest the day before

Ask your recruiter: Which language can I use in each round? Is AI allowed in any round? Is the coding screen CodeSignal levels or live? Is system design blank-page or a doc critique?

Sources

Section titled “Sources”