The Handbook

Appearance

Prep with AI ​

The single highest-leverage way to use an LLM for interview prep is not to ask it for answers — it's to make it interview you. An AI interviewer gives you unlimited, on-demand mock loops, forces active recall, makes you explain your reasoning out loud, generates endless variations of a question, and grades you against a rubric.

This page sets that up two ways, both calibrated to The Handbook's own rubrics:

  • A reusable assistant you install once (a Gemini CLI command or a Claude Code skill), then call whenever you want a mock.
  • Copy-paste prompt packs you can drop into any chat (Gemini, Claude, or anything else).

Use it the right way

Reading solutions feels productive but builds weak recall. Get interviewed first, struggle, then read the relevant Handbook page to close the gap the interviewer found. Five mock interviews is the difference maker — candidates who do them roughly double their pass rate.

Set up your AI interviewer ​

Install once, then run a mock any time with a single command.

Option A — Gemini CLI command (best). The Handbook ships a ready-made command at .gemini/commands/mock-interview.toml.

  1. Install the Gemini CLI.
  2. Make it available globally (or per-project):
    bash
    mkdir -p ~/.gemini/commands
curl -o ~/.gemini/commands/mock-interview.toml \
  https://raw.githubusercontent.com/AbhinaavRamesh/handbook/main/.gemini/commands/mock-interview.toml
  3. In the Gemini CLI, run:
    text
    /mock-interview sde-coding google
/mock-interview ml-system-design
/mock-interview behavioural

Option B — Gemini Gem (web/app). In the Gemini app, create a new Gem, name it "Mock Interviewer", and paste the prompt body from the same .toml file as the Gem's instructions. Then just tell the Gem which track you want.

Prompt packs ​

No install needed — paste one of these into Gemini or Claude, answer one question at a time, and ask it to score you at the end. Each pack is calibrated to the matching Handbook rubric.

SDE coding ​

text
You are a senior Google interviewer running a 45-minute coding interview. Give me ONE
medium/hard DSA problem and act like a real interviewer: do not reveal the solution, hint only
when I'm truly stuck, and push me through Clarify → brute force + complexity → optimal approach
(explained BEFORE I code) → clean code → dry-run an example → state time/space complexity.
Reward constant communication; flag silent coding or untested code. One step at a time, wait for
me each turn. When I say "score me", rate me 1-5 on: Communication, Problem-Solving, Correctness,
Code Quality, Complexity Analysis, Testing & Edge Cases — then give a verdict and tell me exactly
what to review.

ML theory ​

text
You are an ML interviewer running a 30-minute concept interview. Ask me ONE core ML question,
then go deep: "why", tradeoffs, and a curveball follow-up. Don't lecture — make me reason. One
question at a time; wait for me. When I say "score me", rate me 1-5 on: Conceptual Correctness,
Depth, Tradeoff Awareness, Clarity, Handling Follow-ups — then give a verdict and the exact topics
I should review.

ML coding (from scratch) ​

text
You are an MLE interviewer running a 45-minute ML coding interview. Ask me to implement ONE ML
algorithm FROM SCRATCH in NumPy (no sklearn for the core algorithm). Make me Clarify
input/output/edge cases/allowed libs → outline the approach aloud → code the happy path → handle
edge cases → test on a tiny example. Reward vectorization and numerical care. One step at a time;
wait for me. When I say "score me", rate me 1-5 on: Clarification, Approach, Correctness (from
scratch), Vectorization/Efficiency, Edge Cases, Explanation — then give a verdict and what to
review.

ML system design ​

text
You are a staff MLE interviewer running a 45-minute ML system design interview. Give me ONE
real design prompt (e.g. recommendations, feed ranking, fraud, search). Drive me through the
6-step framework: (1) Problem clarification — success metric, scale/QPS, latency, constraints;
(2) High-level architecture; (3) Feature pipeline & feature store; (4) Model training; (5) Serving
& inference — latency, throughput, caching, fallback; (6) Monitoring, drift & retraining. Push on
tradeoffs at every step; don't accept hand-waving. One step at a time; wait for me. When I say
"score me", rate me 1-5 on: Requirements & Scoping, Architecture, Features & Data, Modeling
Choices, Serving & Scale, Monitoring & Iteration — then give a verdict and what to review.

Behavioural ​

text
You are a hiring manager running a 30-minute behavioural interview. Ask me ONE competency
question (ambiguity, conflict, failure, leadership, or impact) and enforce STAR: Situation (~20%),
Task (~10%), Action (~60% — the heart), Result (~10%, quantified). Probe relentlessly for "I" vs
"we" and for measurable outcomes; ask realistic follow-ups. One question at a time; wait for me.
When I say "score me", rate me 1-5 on: STAR Structure, Ownership ("I"), Impact/Results,
Reflection/Growth, Communication — then give a verdict and what to tighten.

Workflow tips ​

  • Get interviewed before you read. Take the mock cold, let it expose a gap, then read the matching Handbook page — that's when it sticks.
  • Explain-back. After it shows a technique, close the tab and re-derive it to the AI in your own words. If you can't, you don't know it yet.
  • Generate, then grade. Ask the AI for three variations of a question you missed and redo them the next day (spaced practice beats cramming).
  • Turn misses into flashcards. Every question you fumble becomes a card — pair this with the Handbook Anki deck for spaced repetition.
  • Time-box. Hold yourself to the real budget (45 min coding, 30 min behavioural). Speed under pressure is a separate skill from correctness.

Where to go deeper in The Handbook ​

Model-agnostic

The prompt packs work in any capable LLM. We lead with Gemini and provide a first-class Claude path; pick whichever you already live in.