Time box: 20 minutes · Format: solo, spoken · What is tested: recall, compression, decision-making
Why this drill exists
Recognition is not recall. Reading about HNSW and finding it familiar feels like knowledge, and it collapses the moment someone asks you to explain it in a sentence. Interviews are a recall task with a social cost: the clock is running, someone is watching, and you cannot scroll back.
Rapid-fire separates the two. It is deliberately too fast for depth, which makes it a clean measurement of what is already available. The gaps it finds are cheap to fix now and expensive to discover in a loop.
Mental model
A flight simulator for recall. You are not learning to fly in the simulator; you are checking which controls you can find without looking down. The value is in the failures it surfaces, not in the hours logged.
The analogy breaks because fluency is not understanding. A fast, confident, wrong answer scores worse than a slow, correct one — the drill rewards compression, and compression can hide a missing mechanism. That is why every answer has to include a reason, and why the follow-up in a real interview exists.
How it works
- 1Pick one question per concept, not per chapter. Twelve questions should span retrieval, evaluation, agents, production and one behavioural prompt. A bank skewed to your strongest area measures nothing.
- 2Answer in a fixed shape. One sentence of claim, one sentence of mechanism, one sentence of when it fails. Sixty seconds is enough for three sentences and not enough for a lecture.
- 3Score immediately, before you feel better about it. Mark each answer as correct, correct-but-empty, or wrong. "Correct but empty" is the important bucket: the claim was right and the reason was missing.
- 4Convert each miss into one card. The card asks the question, not the topic. "Why does reranking not fix missing evidence?" beats "Reranking, chapter 3".
- 5Repeat on a schedule. Re-run misses after two days and again after a week. Re-run the full set monthly; the point is retention, not a single good session.
Step 1 of 5
Pick one question per concept, not per chapter. Twelve questions should span retrieval, evaluation, agents, production and one behavioural prompt. A bank skewed to your strongest area measures nothing.
Figure. The mechanism in this page's own steps. The figure adds sequence, not new claims.
What to notice. A step can be read alone, but the order is the mechanism. The control lets you stop on any one.
Text version
The page's steps: 1. Pick one question per concept, not per chapter. Twelve questions should span retrieval, e… 2. Answer in a fixed shape. One sentence of claim, one sentence of mechanism, one sentence o… 3. Score immediately, before you feel better about it. Mark each answer as correct, correct-… 4. Convert each miss into one card. The card asks the question, not the topic. "Why does rer… 5. Repeat on a schedule. Re-run misses after two days and again after a week. Re-run the ful…
- Pick one question per concept, not per chapter. Twelve questions should span retrieval, evaluation, agents, production and one behavioural prompt. A bank skewed to your strongest area measures nothing.
- Answer in a fixed shape. One sentence of claim, one sentence of mechanism, one sentence of when it fails. Sixty seconds is enough for three sentences and not enough for a lecture.
- Score immediately, before you feel better about it. Mark each answer as correct, correct-but-empty, or wrong. "Correct but empty" is the important bucket: the claim was right and the reason was missing.
- Convert each miss into one card. The card asks the question, not the topic. "Why does reranking not fix missing evidence?" beats "Reranking, chapter 3".
- Repeat on a schedule. Re-run misses after two days and again after a week. Re-run the full set monthly; the point is retention, not a single good session.
The bank
Answers below are the target compression, not a script. If your answer carries the same claim and a reason, it passes even with different words.
| Question | Target answer |
|---|---|
| Why can a reranker not fix missing evidence? | It only reorders retrieved candidates; the passage was never in the list. |
| Cosine similarity versus dot product? | Cosine ignores magnitude, dot product does not; for normalised vectors they rank identically. |
| Why is temperature 0 not deterministic? | Sampling is only one source of variance; batching, hardware kernels and floating-point reduction order also change outputs. |
| What does an embedding dimension buy you? | Capacity to separate concepts, at linear cost in storage and search time; beyond a point, more dimensions add cost without adding separation. |
| Why is HNSW approximate? | It explores a bounded candidate list in a proximity graph, so it can miss a true neighbour that the traversal never visits. |
| TTFT versus total latency? | Time to first token is what the user perceives as responsiveness; total latency is what the request costs and when the answer is complete. |
| Why does chunk overlap exist? | To avoid cutting a fact away from its qualifier at a boundary; too much overlap duplicates evidence and inflates cost. |
| What is the difference between RAG and fine-tuning? | RAG adds evidence at inference time; fine-tuning changes behaviour and format. Neither substitutes for the other. |
| Why is p95 more useful than the mean? | The mean hides the slow tail that users actually complain about. |
| What terminates an agent loop? | An explicit budget or a terminal state from the model; without one, a failing tool produces an unbounded retry. |
| Prompt injection versus jailbreak? | Injection is untrusted content steering the model; a jailbreak is the user trying to bypass policy. Both are instruction-hierarchy failures. |
| Why not use an agent for a fixed pipeline? | A workflow is cheaper, testable and bounded; agency adds variance where the steps are already known. |
Twelve is enough for one session. Rotate the rest in from the curriculum you are weakest on.
Rubric
| Score | What it sounds like | What to do next |
|---|---|---|
| 0 | Silence, or a topic rather than an answer | Re-read the concept, then retry the same question tomorrow |
| 1 | A definition with no reason | Add "because" and "fails when" |
| 2 | Claim and reason, no failure condition | Add one concrete anti-signal |
| 3 | Claim, reason, failure condition, inside a minute | Move to a harder question in the same area |
A 3 is not a perfect answer — it is the correct shape. Depth comes from follow-ups, and the drill deliberately stops before them.
Trade-offs
| Format | Good for | Cost | Fails when |
|---|---|---|---|
| Solo rapid-fire | Recall speed, finding gaps cheaply | No pressure, no follow-ups | You rehearse your own framing and never get corrected |
| Paired rapid-fire | Realism, immediate correction | Requires another person | The partner accepts fluent answers without probing |
| Written quiz | Precision, easy to review later | Slow, no speaking practice | You optimise for the written form an interview does not use |
| Flashcards on misses | Retention between sessions | Fragmented knowledge | Cards test definitions instead of decisions |
| Full mock interview | End-to-end realism and pacing | Expensive, tiring | Used weekly instead of as a checkpoint |
Good for
Recall speed, finding gaps cheaply
Cost
No pressure, no follow-ups
Fails when
You rehearse your own framing and never get corrected
- You rehearse your own framing and never get corrected
- The partner accepts fluent answers without probing
- You optimise for the written form an interview does not use
- Cards test definitions instead of decisions
- Used weekly instead of as a checkpoint
Figure. Each choice with what it buys, what it costs, and the case where it loses.
What to notice. The fourth column is the decision: if the losing case describes your situation, the choice is wrong.
Text version
Four trade-offs: Solo rapid-fire is good for Recall speed, finding gaps cheaply, costs No pressure, no follow-ups, and fails when You rehearse your own framing and never get corrected. Paired rapid-fire is good for Realism, immediate correction, costs Requires another person, and fails when The partner accepts fluent answers without probing. Written quiz is good for Precision, easy to review later, costs Slow, no speaking practice, and fails when You optimise for the written form an interview does not use. Flashcards on misses is good for Retention between sessions, costs Fragmented knowledge, and fails when Cards test definitions instead of decisions. Full mock interview is good for End-to-end realism and pacing, costs Expensive, tiring, and fails when Used weekly instead of as a checkpoint
Failure modes
- Speed without decisions. Answers get shorter and vaguer until they are just labels. Debug: record one minute of audio and count the sentences that contain "because". Fewer than one per answer means you are reciting.
- Practising the strongest topics. The drill feels good and measures nothing. Debug: list the twelve concepts in the bank and mark which ones you have shipped or read twice; if more than half are familiar, replace them.
- No spaced return. One strong session, then nothing for a month. Debug: put the misses in the review queue with a two-day interval and check the retention rate after a week.
- Answers that survive only in your own words. A partner asks "why not the alternative?" and the answer collapses. Debug: for each 3, add the follow-up question you would least like to be asked, and answer that next session.
In production: the real loop
The drill's output is a list, not a feeling. Before an interview, run one session on the twelve questions closest to the role and write down the two answers that were weakest. Rehearse those two in the full six-part shape from the answer framework — the drill gives you the headline, the framework gives you the two minutes behind it.
People think: If I can answer quickly, I understand it.
Actually: Speed measures availability, not depth. The drill is a filter for what needs deeper work; a fast answer still fails when the interviewer asks for the mechanism or the counter-case.
People think: Rapid-fire is only useful for junior candidates.
Actually: Senior loops compress harder, not less: staff-level questions often start with a one-line judgement and spend the rest of the time on consequences. The drill trains exactly that first line.
Interview angle
Interviewers use rapid-fire as a calibration step: three short questions to decide how deep to go. The useful preparation is not memorising answers but practising the compression — a claim, a reason, a failure condition, in that order. If a question lands in an area you have shipped, lead with the concrete case; if it does not, say so and reason from mechanism. See mock interviews, interview formats, and the career interview category.
Sources
- Roediger & Karpicke, Test-enhanced learning — retrieval practice improves long-term retention relative to restudying.
- Dunlosky et al., Improving students' learning with effective learning techniques — practice testing and distributed practice rank as the most effective techniques reviewed.
- Spaced repetition — the scheduling principle behind the two-day and one-week repeats.
- Tech Interview Handbook — question formats and preparation structure for technical loops.