Start from the four rounds
Lesson 1 showed that every loop keeps coming back to four rounds: practical coding, LLM system design, a project deep dive, and motivation and values. The four-round shape is Reported at Anthropic, OpenAI and Perplexity; the exact mix for your role is something only your recruiter can confirm. Your plan should give each round real hours, not just the one you enjoy.
Where this course trains each round
| Round |
What it tests |
Where this course trains it |
| Practical coding |
Building and extending working Python fast, without a mess |
C3 (progressive drills), plus C2 for the LLM facts you'll be asked to code against |
| LLM system design |
Requirements, architecture, evals, failure modes, cost |
C2 (fundamentals), then your track's design module: E6 for engineers, A3 and A6 for architects, the tools, evals and production modules for FDEs |
| Project deep dive |
Defending decisions you made, under follow-ups |
Your own project, or your track's capstone |
| Motivation and values |
Why this lab, how you behave when the right thing is expensive |
C1 (your written answer), C4 (reading the labs), C5 (values and behavioural) |
How long the process takes
Plan your prep against the process, not against a fixed date you picked. At Anthropic, references and team matching after the onsite can add two to four weeks or more (Reported). End to end, third-party guides give anywhere from about three weeks to three months, and they disagree a lot (Anecdotal). A handful of forum accounts say passing the onsite isn't an offer: team matching can run long with little news (Anecdotal). How Anthropic sets levels has no reliable public source; ask your recruiter.
Two practical consequences. First, don't apply before you've started preparing: at Anthropic the online assessment is reported to come straight after the recruiter screen, and you want C3's drills behind you by then. Second, expect a quiet stretch at the end. Silence during team matching isn't necessarily a rejection. Keep your skills warm and your other processes moving.
The plan: six weeks, adjustable to four or eight
About 8–10 hours a week. Keep a small amount of coding practice running every week, because the coding screen usually comes first. Slide the diagram above between four, six and eight weeks to see the plan change.
The six-week plan, with outputs for each week
| Week |
Focus |
Modules |
Output by the end of the week |
| 1 |
The map and your story |
C1 |
Your "Why this lab?" draft (your own first draft, then Claude critique). Your deep-dive project chosen. Questions for the recruiter written down. |
| 2 |
Fundamentals and the format |
C2, C3 lesson 1 |
You can explain tokens, context, tool use, caching and cost math out loud without notes. One untimed drill done. |
| 3 |
Coding under the clock |
C3 drills, timed, tutor off |
Two timed drills completed. A list of the spec details you misread. |
| 4 |
Design |
Your track's design module |
Two full design answers written against the rubric. One talked through out loud. |
| 5 |
The lab and the values round |
C4, C5 |
One written critique of a lab essay. One mock values interview scored. Three STAR stories you can tell in two minutes each. |
| 6 |
Full loop |
Your track's mock loop |
One full mock loop end to end. Fix the weakest round, then re-run just that round. |
- Four weeks? Merge weeks 1–2 and weeks 4–5, and do the mock loop in week 4. Cut depth, not rounds: one drill, one design answer and one values mock are worth more than three of any single one.
- Eight weeks? Add a week of track depth after week 2 (E2–E4 for engineers, A4–A5 for architects, the integrations and agents modules for FDEs), and a second mock loop in week 8.
- Your track shapes weeks 4 and 6. Applied AI Engineers lean into E1–E6 and finish with E7. Architects lean into A1–A6 and finish with A7, where case presentation and customer role-play replace some of the coding. FDEs lean into the customer-engagement modules and use the capstone as their deep-dive project.
You have four weeks and you're strongest at coding. Should you drop the values round to fit in more drills?
No. Each of the four rounds can end your loop on its own, so dropping one is a bad trade, and at Anthropic the culture interview sits in the first onsite loop. Merge weeks and cut depth instead: fewer drills, one design answer, one values mock.
From my loop
Big-company loops tested breadth and process: structured questions, scored rubrics. When I moved to an AI startup as an FDE architect, the interviews tested whether I could work in ambiguity. One round was a live scenario: a customer with a messy problem and half the data, and 45 minutes to scope it, pick an approach, and sketch a working prototype while talking to them.
They were testing whether I'd ask good discovery questions, make a call without perfect information, and own the outcome. To prepare, I ran mock customer calls and practiced narrating tradeoffs out loud while building. If your target role is customer-facing, add that kind of practice to your plan.
Picking your deep-dive project
The project deep dive is Reported at all three labs, and you choose the project, so choose well. Score your candidates:
DEEP-DIVE PROJECT CHECK (one line per candidate project)
Project: ______________________
[ ] Decisions I made myself (I'll be asked "why?" five times in a row)
[ ] A tradeoff I can defend (what I chose, against what, and what it cost)
[ ] Something that went wrong (how I found it, what I changed)
[ ] Real numbers (latency, cost, accuracy, users, hours saved; look them up now)
[ ] Relevance (an LLM system is ideal; a strong non-AI project beats a weak AI one;
customer-facing work is a plus for FDE and architect roles)
[ ] A confidential-safe version (the shape of the work, without the employer's name or code)
If nothing you've built fits, the capstone in your track is designed to. Build it properly and treat it as yours.
Worked example. Candidate A: the team's flagship search system, but a senior colleague made the key design calls. Candidate B: a smaller ticket-triage service where you chose a cheaper model, missed a failure on long tickets, fixed it with an eval, and cut cost per ticket by a number you can name. B wins: it ticks every box, and A fails the first one.
Using recruiter calls to confirm the format
Your recruiter is the only source on your specific loop, and asking good questions makes you look prepared, not anxious. Ask in the first call and again before each stage:
RECRUITER QUESTIONS
- What are the stages for this role, and roughly how long does each take?
- What language and environment is the coding assessment in? One problem or several?
- Is any AI tool allowed in the take-home or the coding rounds?
- What should I prepare for the project deep dive? Slides, or talk it through?
- Is the system design round general, or focused on LLM systems?
- Which level is this role considered for, and how is level decided?
Write the answers down. Where they differ from what you read here or anywhere else, the recruiter wins: that's the most current, most specific source you'll get. Then adjust the plan.
Cheat card
- Six weeks at 8–10 hours a week is the default; four merges weeks (cut depth, not rounds), eight adds track depth and a second mock loop. Coding practice runs every week.
- Pick a deep-dive project with your own decisions, a defensible tradeoff, a failure and real numbers.
- After the onsite, references and team matching can add 2–4+ weeks (reported); end-to-end timelines and levels aren't reliably documented. The recruiter overrides every guide.
Week | Focus | Modules | Output I can show
Interview line "I prepared all four rounds every week, picked a deep-dive project where I made the calls and can show the failure and the numbers, and confirmed the format with my recruiter at each stage."