Get Anthropic interview opportunity, I was honestly a bit stunned — I hadn't had high hopes after applying, so I didn't expect to receive an OA invitation just over a week later.
Anthropic has very little experience in domestic forums. When preparing for the exam, I basically had to rely on sporadic English posts on Glassdoor and Blind. So after the whole process is completed, the first thing I want to do is record it. Regardless of the outcome, I hope this post can help someone in the future.
The process is: OA → Recruiter Screen → 4 rounds of Virtual Onsite, and the whole process takes about three weeks to complete.
Anthropic OA
The platform is CodeSignal, the total duration is 70 minutes, and there are 4 programming questions with increasing difficulty.
CodeSignal's scoring mechanism is different from HackerRank - it doesn't just look at the number of ACs, but scores each question (out of 300/300). Code style, boundary coverage, and operating efficiency will all affect the score. So even if the logic is correct, points will be deducted for messy writing. I didn't notice this at first, but I realized it later.
Question types and directions (recall of real questions):
Q1 (Easy): String processing.
Given a piece of log text, structured fields are parsed according to rules and the formatted results are output. It's basically string segmentation + conditional judgment. There is no algorithmic difficulty, but you should pay attention to edge cases, such as blank lines and missing fields.
Q2 (Easy-Medium): Array operations.
The classic sliding window variant, given an integer array, finds the longest subarray length that satisfies the condition. The condition involves multiple constraints, so you need to understand the logic of window shrinkage.
Q3 (Medium): Graph/tree structure.
Specifically, it is a tree structure problem similar to "employee hierarchical reporting relationship". Given a number of queries, each time the number of nodes that meet the conditions on the path from a node to another node is asked. Examining LCA + DFS, the idea is clear but there are many implementation details.
Q4 (Medium-Hard): Design + Simulation.
To implement a simplified version of "task scheduler" - with task priorities, dependencies, time windows, and output execution order according to rules. This question is the most time-consuming. I was rushing for the last 10 minutes, and there were several cases with conflicting dependencies that I didn't pass completely.
Overall feeling: The time pressure is relatively high, and 70 minutes for 4 questions is not enough for Q3/Q4. It is recommended that Q1/Q2 be controlled within 15 minutes, leaving enough space for the last two questions.
Recruiter Screen (30 minutes)
OA received the email about 5 days later and made an appointment with the recruiter.
The content is relatively relaxed, mainly:
- Introduce Anthropic’s mission and what this team is doing
- Ask me why I am interested in Anthropic and what I know about it
- Confirm timeline, salary expectations, and whether a visa is required
There was a question that made me think about it at the time: "What do you think of AI safety? Do you think there is a contradiction between it and product development?" Anthropic is very concerned about whether candidates really agree with their mission. They don't just use this as a springboard for a fancy AI company. It is recommended to carefully prepare for such questions.
Virtual Onsite (4 rounds, same day)
Do them consecutively on the same day, with a 10-minute break in between. Each round lasts 45 to 60 minutes.
Round 1: Coding
The interviewer was from the Core Infrastructure team. He was very nice and we chatted for a few minutes.
Topic:Design a Rate Limiter.
It is not a pure algorithm question, but from the perspective of functional implementation: first implement a basic version (fixed window), and then the interviewer asks whether it can be changed to a sliding window, then asks how to expand if it wants to support multiple users, and finally talks about the challenges in a distributed scenario.
I implemented it in Python. The whole process was written while speaking. The interviewer would interrupt at the right time and ask "Why did you choose this way?" It feels like he cares more about whether your thinking is clear rather than whether there are bugs in the code.
Hint: Rate Limiter is a high-frequency Anthropic question. You must be familiar with the advantages and disadvantages of the fixed window, sliding window, and Token Bucket solutions in advance.
Round 2: Coding
The topic is more about business logic, probably: Implementing a simplified version of the conversation context manager.
Given a token upper limit, the conversation history is maintained. When new messages are added and the token limit is exceeded, old messages need to be pruned according to the policy. The strategy is divided into several follow-up iterations: first simple FIFO clipping, then changing to retaining the system prompt without clipping, and finally adding priority labels.
This question is obviously directly related to Claude's actual product scenario. When I was doing it, I felt like I was writing infrastructure for Claude, which was a bit strange. The amount of code is not large, but you need to think clearly about boundary situations, such as how to handle a single message that exceeds the token limit.
Round 3: Behavioral
This round is the most important round for me to prepare for. The interviewer is the Engineering Manager.
High frequency problem recall:
- "Tell me about a time when you made a technical decision that you thought was correct but the team didn't agree with it. How did you resolve it in the end?"
- “What’s the most complex trade-off you’ve ever done on a project?”
- “Have you ever proactively identified an important problem that no one was following but pushed for a solution?”
- “How do you tell whether a technical solution is good enough to ship?”
The behavioral aspect of Anthropic places great emphasis on initiative and judgment, not just "I completed the task", but "Why did I judge this way? I saw something that others did not see." The STAR framework can be used, but don’t be too templated and talk more about your own reasoning.
Round 4: Values & Mission Fit
Many people are not prepared for this round, but it is actually quite important. The interviewer was a Senior Researcher and the conversation was very in-depth.
Core topics:
- What do you understand by AI safety? Do you think Anthropic is heading in the right direction?
- “What would you do if you were developing a feature and you discovered that it might be at risk of abuse?”
- “What do you think about the tension between helpful and harmless?”
There are no standard answers in this round. What the interviewer wants to see is whether you have really thought about these questions independently, rather than memorizing an introduction from the official website. I mentioned that I had read the Constitutional AI paper published by Anthropic, and the interviewer immediately started chatting a lot, and the atmosphere was very good.
Suggestion: Read Anthropic’s research blog, Claude’s model card, and their public AI safety-related articles in advance. Having your own opinions is more important than anything else.
Exam preparation checklist
- Question brushing: Mainly LeetCode Medium, focus on graph/tree/sliding window/design questions.
- System design: Understand the basic principles of common current limiting, caching, and message queues
- Behavioral questions: Prepare 5 to 6 in-depth STAR stories, covering the three directions of conflict, trade-off, and initiative
- AI cognition: read Anthropic blog, understand Constitutional AI, and have your own views on safety vs capability
A little extra | About exam preparation efficiency
The above is the process of chewing it off myself. But to be honest, Anthropic has too little interview experience, and the question style of OA is also different from LeetCode - it is more realistic and the boundary conditions are more hidden. It is even more critical when it comes to the VO stage. Anthropic's interviewers will ask in-depth questions about your ideas, complexity analysis, and even the consistency of your coding style. To be on the safe side, I made a reservation in advance. ProgramHelp VO real-time assist service is equivalent to giving yourself a layer of insurance.
If you also face the following situation:
- Face-to-face experience is scarce, so I don’t know what I’m currently taking the exam for.
- OA has ideas, but always falls over on hidden cases.
- VO is worried about getting stuck on site, not being able to express his ideas clearly enough, and not being able to keep up with the interviewer’s questioning.
You might as well prepare an extra trump card for yourself, and your mentality will be much more stable when you go on the court.
