design-room

A system-design interviewer that will not help you.

Docs License: MIT 62 questions Runs on Claude Code

design-room is a Claude-Code-driven mock interview for senior, staff and Engineering-Manager system-design loops. Claude plays a FAANG interviewer, holds the clock, probes your decisions, and scores you against a nine-dimension rubric. You design in interview.md and talk to it in chat.

It is not a chatbot that explains system design to you. It is a room with a stranger in it who will not help, will not praise you, and will stop on time. That is the point — the failure modes it trains away (not scoping, not managing the clock, not going deep, not talking about operations or cost) are behavioural, and they only surface under those conditions.

1 · Start a session

You need Claude Code. Nothing else — no install, no dependencies, no API key beyond the one Claude Code already uses.

git clone https://github.com/nirmata78/design-room.git
cd design-room
claude

Then say:

Run a system design mock.

That is the whole invocation. CLAUDE.md at the repository root is loaded automatically: it is the interviewer’s operating manual. Claude asks you the six setup questions once, creates interview.md from the template, starts a real clock, and reads you a problem.

To skip the setup questions, specify what you want:

Mock me: SD-08, Google, L6, full round, brutal.
Surprise me — pick my weakest archetype.
15-minute framing drill, three questions back to back.

Optional — tell it who you are. Copy profile.example.md to profile.md and fill it in: target companies and level, your home domain (it probes you harder there), and any confidentiality rules about your employer. profile.md is gitignored and never leaves your machine.

No Claude Code? prompts/ has a paste-into-any-chat version. It has no clock, no workspace and no memory between rounds, so it is a much weaker simulation — but it works anywhere.

2 · How a round actually goes 

   YOU                                        CLAUDE (interviewer)
   ───                                        ────────────────────
                                              reads the prompt aloud, once
                                              starts the clock (real: `date`)
   write §1 Requirements in interview.md
   "requirements are up"            ────▶     reads §1 only, asks ONE probe
   answer in chat                              ◀──── "what did you deliberately leave out?"
   write §2 Estimation
   "done"                          ────▶      reads §2, probes the number
   ...                                        moves you on when a phase overruns
   write §4 the diagram
   "take a look"                   ────▶      "walk me through one request end to end"
   ...                                        deep dives: 3–4 probes, one curveball
   §8 four-sentence summary        ────▶      stops on time
                                              ─── drops the persona ───
   "how did that go?"              ◀────      asks YOU first, then the debrief
                                              writes the scorecard, archives the session

Your side of the contract: write in §1–§9 of interview.md, and say something in chat when you want it looked at. Rough and fast beats neat and slow — nobody scores your ASCII art. Most of your reasoning happens in chat, not in the file.

Its side: it reads only the sections the current phase covers, writes only to §10, asks one question at a time, and does not help you.

What it will not do

It won’t name a component you haven’t named, tell you whether an answer is right, say “good”, complete your diagram, or fill a silence. If you ask “should I use Kafka here?” you’ll get “which would you pick, and why?” All the teaching is held back for the debrief, where it’s worth ten times more.

If you want it to teach mid-round, say coach mode — it breaks character, the clock stops, and the round stops being scoreable. That’s a deliberate trade, not a failure.

3 · Talking to it

Say What happens
done · ready · look at §2 Gate: it checks the clock, reads that section, probes
time? Elapsed and remaining in this phase. Nothing else
stuck · hint The hint ladder, starting at the weakest rung — and it records which rung you needed
repeat the prompt Verbatim, no additions
assume 50M users Accepted and written into §0 as an established fact for the round
harder · easier Shifts the difficulty dial mid-round
curveball Ask for one now, if you want the practice
skip estimation Allowed — and it’ll cost you at debrief
pause / resume Stops and restarts the clock, excluding the gap
debrief · stop Ends the round now and goes straight to feedback
coach mode Breaks character to teach. Round becomes unscoreable

4 · The modes

Mode Time Use it when
full 75 min The real thing. Full scorecard. Default
framing 15 min The best value-per-minute drill. Three prompts, scope + estimate only, no design. Trains the #1 failure mode
hld 30 min Through the diagram, stop before deep dives
deep-dive untimed One component, drilled to the bone
grill untimed Adversarial probing on a design you already have. 8+ probes, no mercy
wrap 12 min Failure modes, SLOs, cost and the four-sentence summary — the EM-bar phase alone
review 20 min Inverted: Claude presents a design, you critique it. Tests the skill an EM uses weekly
loop — Two or three rounds back to back with a cross-round pattern note

If you have twenty minutes, run framing. If you have one hour, run full. If you have a loop next week, run loop with the company’s flagship questions.

5 · The question bank — 62 questions

Question-Bank.md is the index. Eight archetypes, 29 with full cue sheets (crux, clarification numbers, checklist, phase-by-phase probes, curveball, weak/strong/exceptional, traps) and 33 compact ones, all runnable.

  Archetype Qs The recurring test
A Ingest, streaming & real-time analytics 7 Durable buffer, event time, approximate vs exact
B Social, feed & content 7 Fan-out — pay at write or read, and the outlier
C Storage, data & consistency 9 Honesty about consistency
D Infrastructure & control planes 12 Control plane vs data plane
E Search, geo & ML 8 Precomputation under a latency budget
F Transactions, payments & commerce 7 Correctness under retry and partial failure
G Media & delivery 4 Moving bytes cheaply — egress dominates
H EM-specific rounds 8 Reviewing, sequencing, budgeting, SLOs, splitting work

Archetype H is the one to not skip. It holds the rounds that aren’t “design X”: critique someone else’s design, sequence an eighteen-month migration, take 30% out of an infrastructure bill, fix a service with 1,200 alerts and an exhausted team, split a system across five teams. Four of them are inverted — Claude presents an artifact (a design doc, a bill, an incident timeline) with planted defects, and you review it. An L6 EM loop very often contains one of these, sometimes disguised as a normal design round that turns into a review halfway through.

Company flagships, if a loop is close:

Company Run these first
Google L6 EM SD-15 global KV · SD-36 crawler · SD-02 top-K · SD-25 rate limiter · SD-19 collaborative editing · SD-32 CI/CD · SD-58 SLOs · SD-59 team split
Meta E6 EM SD-08 feed · SD-09 messaging · SD-20 cache · SD-41 recommendations · SD-10 live comments
Amazon L6/L7 SD-48 inventory · SD-44 ledger · SD-47 booking · SD-28 workflows · SD-56 migration · SD-57 cost · SD-61 incident
Microsoft SD-23 multi-tenancy · SD-19 collaborative editing · SD-09 messaging · SD-34 secrets · SD-59 team split
Netflix SD-51 video · SD-52 live · SD-41 recommendations · SD-35 rollout · SD-57 cost

No loop yet? Go for coverage — one round in each of A–H teaches more than eight rounds in the two archetypes that feel comfortable.

6 · Scoring

Nine dimensions, 1–4. 1 below bar · 2 approaching · 3 at bar (hire) · 4 above bar. A strong round is mostly 3s with a couple of 2s and 4s.

  1. Requirements & scoping
  2. Estimation — did a number change a decision
  3. High-level design
  4. Technical depth — two levels deep when probed
  5. Trade-off reasoning
  6. Operations, reliability & cost — the manager bar
  7. Communication & collaboration
  8. Handling ambiguity & pushback
  9. Leadership signal (EM)

Plus three things that often say more than the scores: which hint rungs you needed, where the time actually went, and how accurate your own self-assessment was.

Every score comes with a quote or a timestamp. A dimension it never probed is left unscored, not given a default. And it audits itself at the bottom of the sheet — leaks, missed probes, whether it under-probed — because an inflated score is the one way this thing can actively hurt you.

Rubric: Scorecard-Template.md. It mirrors guides/interviewer-guide.md §7 — a guide written for a human mock partner, even a non-technical one, with a printable grading sheet — so a peer running you through it on paper scores you on the same axes.

7 · Files 

design-room/
├── CLAUDE.md                the interviewer's operating manual — the engine. Read by Claude, not you
├── README.md                this file
├── Question-Bank.md         index of all 62 questions
├── questions/
│   ├── A-streaming-analytics.md           SD-01…07
│   ├── B-social-feed-content.md           SD-08…14
│   ├── C-storage-data-consistency.md      SD-15…23
│   ├── D-infrastructure-control-planes.md SD-24…35
│   ├── E-search-geo-ml.md                 SD-36…43
│   ├── F-transactions-payments.md         SD-44…50
│   ├── G-media-delivery.md                SD-51…54
│   └── H-em-specific-rounds.md            SD-55…62  ← includes the inverted rounds + answer keys
├── templates/
│   └── interview-template.md   copied to ./interview.md at session start
├── Scorecard-Template.md    the nine-dimension rubric
├── Cue-Sheet-Template.md    for running a prompt that isn't in the bank
├── Progress-Log.template.md copied to Progress-Log.md on first run
├── profile.example.md       copy to profile.md to calibrate the interviewer to you
├── guides/
│   ├── interviewer-guide.md the persona — also usable by a human mock partner
│   └── playbook.md          how to behave in the room; estimation number kit; phrasebook
├── prompts/                 paste-into-any-chat fallbacks (system design, behavioural, coding)
│
│   ── private, gitignored ──
├── profile.md               who you are and what you're targeting
├── Progress-Log.md          one row per round; drives "surprise me" and coverage tracking
├── interview.md             YOUR WORKSPACE — exists only during a session
└── sessions/                archived rounds: interview.md + scorecard.md + debrief.md

Your rounds stay private. profile.md, Progress-Log.md, interview.md and everything under sessions/ are in .gitignore. If you fork this repository, your scores and designs will not be pushed with it.

Don’t read questions/*.md before a round. They contain the answer — the checklist, the probe ladder, the model design, and for archetype H the planted-defect keys. Reading the sheet turns a mock into a recital. Read them after, at debrief, when Claude walks you through what you missed.

8 · The loop that builds skill

  1. Run a round. Any mode. Don’t pick the question yourself once you’ve done a few — surprise me reads Progress-Log.md and picks your weakest archetype.
  2. Take the debrief seriously, especially step 1 where it asks how you think it went before saying anything. Being able to grade yourself accurately is most of what separates people who improve from people who plateau.
  3. Do the one drill. Exactly one thing, small enough to do this week. It’s logged; if the same drill appears three times undone, it was too big — ask for a smaller one.
  4. Study the gap, not the question. If depth on consistency was the weakness, read about consistency — not the model answer to the question you just did.
  5. Re-run cold, weeks later. Not the next day.

Cadence that works four weeks out: week 1 coverage (one round per archetype, A–H) · week 2 depth (re-run the two weakest, deep-dive mode) · week 3 reps (daily framing drill + two full rounds at brutal) · week 4 full loops with company calibration.

If you have one week: the framing drill daily, plus two full rounds, plus one archetype-H round.

9 · Two things worth knowing

Do it out loud where you can. Typing a design and narrating one are different skills, and only the second one transfers to the room. pace: realtime with the spoken answer in chat via voice is the closest simulation; the default typing-adjusted (75 min instead of 50) exists so you’re not penalised for typing speed when working in the file.

Your home turf is a risk. If profile.md names a domain you work in professionally, the interviewer probes you harder there and watches for the expert’s failure mode: assuming the interviewer shares your context and under-explaining. In a real loop the interviewer probably won’t be from your domain, and an expert who can’t make a non-expert follow them scores badly on Communication. Claude will ask you to define your own vocabulary. That’s deliberate.

Also — session files persist. Nothing internal, confidential or identifiable goes into them; Claude is instructed to generalise (“a large enterprise customer portfolio”) rather than record real system names, figures or ticket IDs — and the files are gitignored regardless.

10 · Contributing

New questions, sharper probes and better company calibration are all welcome — see CONTRIBUTING.md. The cue-sheet format is in Cue-Sheet-Template.md.

License

MIT.