interview-reps 0.4.0 → 0.6.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (50) hide show
  1. package/README.md +6 -0
  2. package/content/CLAUDE.md +54 -8
  3. package/content/assessments/A01-library-desk/README.md +13 -0
  4. package/content/assessments/A01-library-desk/level-1.md +17 -0
  5. package/content/assessments/A01-library-desk/level-2.md +16 -0
  6. package/content/assessments/A01-library-desk/level-3.md +16 -0
  7. package/content/assessments/A01-library-desk/level-4.md +15 -0
  8. package/content/assessments/A02-ticket-counter/README.md +16 -0
  9. package/content/assessments/A02-ticket-counter/level-1.md +18 -0
  10. package/content/assessments/A02-ticket-counter/level-2.md +16 -0
  11. package/content/assessments/A02-ticket-counter/level-3.md +16 -0
  12. package/content/assessments/A02-ticket-counter/level-4.md +16 -0
  13. package/content/assessments/REVIEW.md +70 -0
  14. package/content/behavioral/QUESTIONS.md +44 -0
  15. package/content/behavioral/REVIEW.md +92 -4
  16. package/content/design/D13-video-streaming/README.md +21 -0
  17. package/content/design/D14-ticket-booking/README.md +20 -0
  18. package/content/design/D15-ad-click-aggregator/README.md +20 -0
  19. package/content/design/D16-distributed-key-value-store/README.md +21 -0
  20. package/content/design/D17-collaborative-document-editor/README.md +20 -0
  21. package/content/design/D18-ai-assistant-on-a-budget/README.md +21 -0
  22. package/content/design/D19-going-multi-region/README.md +20 -0
  23. package/content/design/D20-checkout-at-10x-load/README.md +20 -0
  24. package/content/design/D21-webhook-delivery/README.md +20 -0
  25. package/content/design/D22-game-leaderboard/README.md +20 -0
  26. package/content/design/D23-discovery-recommendations/README.md +21 -0
  27. package/content/design/D24-splitting-a-shared-database/README.md +20 -0
  28. package/content/design/REVIEW.md +70 -0
  29. package/content/product/P21-support-reply-drafts/README.md +29 -0
  30. package/content/product/P22-receipt-categories/README.md +29 -0
  31. package/content/product/P23-pharmacy-symptom-helper/README.md +31 -0
  32. package/content/product/P24-listing-descriptions/README.md +29 -0
  33. package/content/product/P25-group-class-bookings/README.md +28 -0
  34. package/content/product/P26-home-repair-press-release/README.md +28 -0
  35. package/content/product/P27-merchant-cash-advances/README.md +30 -0
  36. package/content/product/P28-retire-the-desktop-app/README.md +28 -0
  37. package/content/product/P29-live-courier-location/README.md +28 -0
  38. package/content/product/P30-carts-across-devices/README.md +28 -0
  39. package/content/product/P31-ticket-on-sale/README.md +29 -0
  40. package/content/product/P32-transfer-status/README.md +28 -0
  41. package/content/product/P33-new-revenue-by-plan/README.md +47 -0
  42. package/content/product/P34-weekly-app-opens/README.md +46 -0
  43. package/content/product/P35-second-orders/README.md +49 -0
  44. package/content/product/P36-on-time-by-city/README.md +46 -0
  45. package/content/product/P37-trials-that-never-pay/README.md +52 -0
  46. package/content/product/P38-smaller-grocery-baskets/README.md +51 -0
  47. package/content/product/P39-clinic-bookings-abroad/README.md +52 -0
  48. package/content/product/REVIEW.md +136 -1
  49. package/dist/cli.mjs +3615 -797
  50. package/package.json +1 -1
@@ -49,15 +49,39 @@ Give each a one-line reason that points at the answer itself.
49
49
  Also check that the story actually answers the question and shows its competency. A "conflict" story with
50
50
  no disagreement in it does not.
51
51
 
52
+ ## Level bar (when the app gives a target)
53
+
54
+ Some answers come with a target, such as "staff engineer", "senior product manager" or "engineering manager".
55
+ Then also judge whether the story reaches that level's scope. Without a target, skip this section and the
56
+ Level shown line.
57
+
58
+ | Target | Scope a 4 on impact needs | Ownership looks like |
59
+ |---|---|---|
60
+ | Junior, mid | Their own task, done well, with a result for users or the team | A clear "I": what they decided inside the task |
61
+ | Senior | A project they led end to end; a result for the team's product, ideally measured | Trade-offs they weighed and chose; people they brought along |
62
+ | Staff | Beyond their team: other teams or the organisation changed what they build or how | They set a direction others followed, without authority |
63
+ | Principal | Company-wide or over several years; a bet that shaped strategy | They framed the problem that leadership then funded |
64
+ | Engineering manager | Results through people: delivery, hiring, growth, keeping people | Calls they made about people and priorities, and what they stopped |
65
+ | Product manager | Senior: one product area and its main metric. Staff or above: strategy across areas | The decision, the evidence for it, and what they said no to |
66
+
67
+ - If an otherwise good story falls below the target's scope, cap **impact** (or **ownership**, whichever
68
+ falls short) at 3, add `behavioral:scope`, and say in a "Level bar" part of the review which part of the
69
+ scope is missing.
70
+ - Never suggest a bigger story to tell or invent scope. Name the kind of gap, as everywhere else.
71
+ - Print the **Level shown** header line: the level the story actually shows (`junior`, `mid`, `senior`,
72
+ `staff`, `principal` or `manager`), then optionally ` — ` and a reason of at most 120 characters that
73
+ names the kind of gap without retelling the story.
74
+
52
75
  ## Verdict, gaps and Next time
53
76
 
54
77
  - **Verdict:** `solid` only if no dimension is below 3. Otherwise `redo`.
55
78
  - **Gaps:** use only these tags, comma-separated, or `none`:
56
79
  `behavioral:structure`, `behavioral:specificity`, `behavioral:impact`, `behavioral:ownership`,
57
80
  `behavioral:brevity`, `behavioral:self-awareness`, `behavioral:conflict`, `behavioral:failure`,
58
- `behavioral:influence`, `communication`. Add a dimension's tag when it scored below 3. Use
81
+ `behavioral:influence`, `behavioral:scope`, `communication`. Add a dimension's tag when it scored below 3. Use
59
82
  `behavioral:conflict`, `behavioral:failure` or `behavioral:influence` when a question on that
60
- competency got a story that does not show it. Use `communication` for unclear wording or jargon.
83
+ competency got a story that does not show it. Use `behavioral:scope` when the story is below the
84
+ target's level (see Level bar). Use `communication` for unclear wording or jargon.
61
85
  - **Next time:** one concrete habit for the next answer (for example "open the Result with a number"),
62
86
  not a list, and not a rewritten answer.
63
87
 
@@ -72,6 +96,7 @@ Date: YYYY-MM-DD
72
96
  Verdict: solid | redo
73
97
  Scores: structure <1-4>, specificity <1-4>, impact <1-4>, ownership <1-4>, brevity <1-4>, self-awareness <1-4>
74
98
  Gaps: <comma-separated tags, or none>
99
+ Level shown: <junior | mid | senior | staff | principal | manager> — <optional reason> (only with a target)
75
100
  Next time: <one concrete habit>
76
101
  ```
77
102
 
@@ -81,6 +106,69 @@ Then, briefly:
81
106
  2. **Does it answer the question?** One or two lines.
82
107
  3. **Did last time's feedback stick?** Compare with the earlier reviews' gaps and Next time lines, or say
83
108
  this is the first review.
84
- 4. **What an interviewer would ask next**: one or two follow-up questions they should be ready for.
85
- 5. **Communication: pacing and fillers** (spoken answers only): words per minute against the easy range,
109
+ 4. **Level bar** (only with a target), in this form, which the app shows as a block:
110
+
111
+ ```
112
+ ## Level bar · <target, for example staff engineer>
113
+ - Shown: <a part of the target's scope the story shows>
114
+ - Not shown: <a part it does not show>
115
+ Capped: impact
116
+ <one or two sentences on which part is missing>
117
+ ```
118
+
119
+ Use two to four `Shown` or `Not shown` lines taken from the target's row in the Level bar table. Add the
120
+ `Capped:` line (`impact` or `ownership`) only when you capped a score for scope.
121
+ 5. **What an interviewer would ask next**: one or two follow-up questions they should be ready for.
122
+ 6. **Communication: pacing and fillers** (spoken answers only): words per minute against the easy range,
86
123
  the fillers that stood out, and one thing to try. Skip this for typed answers.
124
+
125
+ ## Teaching the review
126
+
127
+ The hub asks for a teaching review: it sends the person's save points (drafts of the answer), how far
128
+ they went in the Help tab, and their finish answers. Teach, don't only state findings. Use these exact
129
+ `##` headings, so the hub can lay the review out.
130
+
131
+ **Open** with one or two plain sentences on how it went: how the story was built, and what cost the
132
+ time.
133
+
134
+ **`## Between your drafts`**, right after the opening, only when there are two or more save points.
135
+ Name the drafts where something changed (for example `### Drafts 1 to 2: the Result gained a number`),
136
+ quote only the changed lines (at most 12 lines in one block), and add **What it means:** in one to three
137
+ sentences on what the change shows about how they worked. Leave the section out with one draft or none.
138
+ The story stays private: quote only the changed lines you need.
139
+
140
+ **Plan against what they finished.** After point 2 (Does it answer the question?), say in a line or
141
+ two whether the answer they finished kept to the story and point they set out to make. When they
142
+ answered "What changed from your plan", use their words.
143
+
144
+ **`## What to work on, with examples`**, after point 3. One `###` entry per finding that has a gap tag,
145
+ written out in full:
146
+
147
+ - **What we saw:** a few of their words, quoted, or what the drafts showed.
148
+ - **Why it matters:** one to three sentences: what goes wrong, and what an interviewer makes of it.
149
+ - **Before** and **After**: the same short fragment (at most 6 lines each) of a different, made-up
150
+ story, one or two sentences, for example about a release checklist at a made-up company called
151
+ Northwind Labs, first told the weak way, then the better way.
152
+ - **How to practise:** a short drill in words, or another question from the bank by its id.
153
+ - **What to say:** one sentence they could say in the interview.
154
+
155
+ The fragment teaches the habit (a result with a size, "I" for their own decisions, a lesson they acted
156
+ on) in someone else's made-up story. It never retells or improves their story, never suggests a story
157
+ to tell, and never names the detail that would fix this answer, so the rule at the top still holds.
158
+
159
+ **What I learned.** If they wrote one, say in one line whether it names the main gap.
160
+
161
+ At the end, after the last numbered point, three short sections the hub shows beside the review, in
162
+ this order:
163
+
164
+ 1. **`## How you solved it`**: a list with Result (*solved alone*; *solved with questions* when help
165
+ stopped at step 4 or earlier; *solved with help* from step 5 on), Help (the furthest step and its
166
+ name, or none), Moved you (only when help was used: the step that got them moving), Time and
167
+ Drafts. Add one line on why it counts as that result.
168
+ 2. **`## Gaps noted, and why`**: one line per tag in the Gaps header line: why this tag, in plain
169
+ words, and what "fixed" looks like. Then `Not tagged:` with a tag you considered and left out, and
170
+ why. Leave this section out when Gaps is none.
171
+ 3. **`## When it comes back`**: `- Redo:` a weekday and date, two or three days out (three when solved
172
+ with help; for solid, a look back in about a week), and how: from a blank page, no notes, no help,
173
+ within about two to three minutes. `- Then:` a sibling question from the bank, by its id, on the same
174
+ competency. `- Daily:` a two or three minute drill for the main gap.
@@ -0,0 +1,21 @@
1
+ # D13 · Video streaming
2
+
3
+ Category: Media · Level: core · Time: 45 min
4
+
5
+ Build a video service where creators upload videos and viewers anywhere in the world watch them on phones, laptops and TVs. Playback should start quickly and keep going without pausing, even on a slow or patchy connection.
6
+
7
+ Before you design, find out how many videos are uploaded per day and how long they are, how many hours are watched per day, which devices and connection speeds must be supported, how soon after upload a video must be watchable, and how evenly views are spread across videos.
8
+
9
+ ## If you ask
10
+
11
+ - About 500,000 new videos a day, averaging 8 minutes; uploads can be up to 10 GB.
12
+ - 1 billion hours watched per day. A small share of videos gets most of the views, and a new hit can go from zero to millions of views in an hour.
13
+ - Every video must play on everything from a 3G phone to a 4K TV, and playback should start within 2 seconds.
14
+ - A new upload should be watchable within 10 minutes. View counts may lag by a few minutes.
15
+ - Out of scope: live streaming, recommendations, comments and ads.
16
+
17
+ ## Follow-ups
18
+
19
+ - Viewers in one country are cut off from your nearest servers for a day. What do they experience, and what changes?
20
+ - Creators now want to stream live, with viewers no more than 5 seconds behind. What changes?
21
+ - Storage costs are growing faster than revenue. What would you change about what is kept, and for how long?
@@ -0,0 +1,20 @@
1
+ # D14 · Ticket booking
2
+
3
+ Category: Marketplace · Level: core · Time: 45 min
4
+
5
+ Design the booking system for concert and sports tickets. Fans look at a seating map, pick seats, and hold them for a few minutes while they pay. A seat must never be sold twice, and a seat whose hold runs out goes back on sale.
6
+
7
+ Before you design, find out how many events and seats there are, how much traffic arrives when a popular event goes on sale, how long a hold lasts, what happens when payment fails or is slow, and whether fans can choose exact seats or only a section.
8
+
9
+ ## If you ask
10
+
11
+ - About 50,000 events on sale at any time, from small clubs to stadiums of 80,000 seats.
12
+ - When a big tour goes on sale, up to 2 million people try to buy the same 20,000 seats within the first ten minutes.
13
+ - A hold lasts 8 minutes. Up to 6 seats per order, and the seats in one order are all held or none are.
14
+ - The seating map may be a few seconds out of date, but a fan must never pay for a seat that someone else has.
15
+ - Out of scope: reselling between fans, pricing, and choosing who gets in first when demand is high.
16
+
17
+ ## Follow-ups
18
+
19
+ - Bots grab holds on thousands of seats and let them expire, again and again. What changes?
20
+ - The same system now sells hotel rooms, where one booking holds a room for several nights in a row. What changes?
@@ -0,0 +1,20 @@
1
+ # D15 · Ad click aggregator
2
+
3
+ Category: Data · Level: core · Time: 45 min
4
+
5
+ Build the system that counts clicks on ads. Advertisers watch clicks per ad per minute on a dashboard, and at the end of each day they are billed per click, so the billed numbers must be right and checkable later.
6
+
7
+ Before you design, find out how many clicks arrive per second at peak, how quickly the dashboard must reflect a click, which breakdowns advertisers want, how long the data is kept, and what counts as a click that should not be billed.
8
+
9
+ ## If you ask
10
+
11
+ - About 10 billion clicks a day, peaking at 500,000 per second; 2 million active ads.
12
+ - The dashboard may be up to 1 minute behind. Billing totals are final once a day, and must match the raw clicks exactly.
13
+ - Advertisers filter by ad, campaign, country and device, at minute detail for 30 days and day detail for 2 years.
14
+ - Some clicks arrive late, up to an hour after they happened, from phones that were offline. The same click can also be sent twice.
15
+ - Out of scope: deciding which ad to show, and fraud scoring (another team flags fraudulent clicks after the fact).
16
+
17
+ ## Follow-ups
18
+
19
+ - A bug double-counted every click from one region for three hours yesterday, and bills have already gone out. What happens next?
20
+ - Advertisers now want to see how many distinct people clicked, not just how many clicks there were. What changes?
@@ -0,0 +1,21 @@
1
+ # D16 · Distributed key-value store
2
+
3
+ Category: Storage · Level: advanced · Time: 45 min
4
+
5
+ Design the storage system itself: a store that other teams use to read and write values by key, spread across hundreds of machines. It keeps working when machines fail, and each team chooses how up to date its reads need to be.
6
+
7
+ Before you design, find out how much data is stored and how large values are, how many reads and writes per second there are and in what mix, what happens when two writes to the same key race, how much downtime is acceptable, and whether data must survive the loss of a whole rack or data centre.
8
+
9
+ ## If you ask
10
+
11
+ - About 500 TB in total, with keys under 100 bytes and values usually under 10 KB (never more than 1 MB).
12
+ - 2 million reads and 500,000 writes per second at peak, with a few keys far busier than the rest.
13
+ - The store runs in one region across three data centres, and must keep accepting reads and writes if one data centre is lost.
14
+ - Some teams need to read their own latest write every time; others accept reads that are a few seconds stale in exchange for speed.
15
+ - Out of scope: queries other than by key, transactions across several keys, and multiple regions.
16
+
17
+ ## Follow-ups
18
+
19
+ - You need to add 50% more machines without slowing anyone down. Walk through what happens.
20
+ - A team wants to run it purely as a cache in front of their database, where losing data is fine but speed matters most. What would you change?
21
+ - Two data centres can no longer talk to each other, but clients can reach both. What do reads and writes return?
@@ -0,0 +1,20 @@
1
+ # D17 · Collaborative document editor
2
+
3
+ Category: Web · Level: advanced · Time: 45 min
4
+
5
+ Design an online document editor where several people type in the same document at the same time. Each person sees the others' changes and cursors almost immediately, can keep editing offline, and nobody's text is lost when the edits meet.
6
+
7
+ Before you design, find out how many documents and editors there are, how many people edit one document at once, how fast others must see a change, how long people work offline, what history must be kept, and who can view or edit a document.
8
+
9
+ ## If you ask
10
+
11
+ - 50 million daily users and 1 billion documents; most documents are under 100 KB of text.
12
+ - Usually 1 to 5 people edit a document at once, but a shared meeting document can have 200.
13
+ - Others should see a change within half a second. People can work offline for days and then reconnect.
14
+ - Everyone must end up seeing exactly the same document, and people can browse and restore earlier versions.
15
+ - Out of scope: images and tables inside documents, comments, and spell checking.
16
+
17
+ ## Follow-ups
18
+
19
+ - A company-wide announcement document is opened by 50,000 people at once while three people edit it. What changes?
20
+ - Someone reconnects after two weeks offline, having rewritten a section that others deleted. What do they see?
@@ -0,0 +1,21 @@
1
+ # D18 · AI assistant on a budget
2
+
3
+ Category: ML · Level: advanced · Time: 45 min
4
+
5
+ A customer support product adds a button that summarises a long ticket thread and drafts a reply, using a large language model. Design the system behind the button. Spending on the model has a hard monthly ceiling, answers must arrive quickly, and one large customer must not crowd out everyone else.
6
+
7
+ Before you design, find out how many companies and agents use the product, how often the button is pressed and when, how long threads are, the monthly budget and what one model call costs, how fast an answer must start appearing, and what happens when the budget or the model provider runs out.
8
+
9
+ ## If you ask
10
+
11
+ - 50,000 customer companies with 2 million support agents; the button is pressed about 20 million times a day, mostly during working hours in each time zone.
12
+ - A thread averages 3,000 words and the longest are 50,000. A reply is about 200 words.
13
+ - The monthly budget is fixed at an amount that covers only about 60% of today's presses at the price of the largest model. There are also smaller models that cost a tenth as much and are less accurate.
14
+ - The first words should appear within 2 seconds. Customers' data must never appear in another customer's answer.
15
+ - Out of scope: training or fine-tuning models, and the wording of the prompt sent to the model.
16
+
17
+ ## Follow-ups
18
+
19
+ - The model provider halves its rate limit for your account for a day. What do agents see?
20
+ - Product wants to know whether the drafts are actually good. How would you measure that, and what does the system need to record?
21
+ - The ten largest customers want a guaranteed share of capacity, and will pay for it. What changes?
@@ -0,0 +1,20 @@
1
+ # D19 · Going multi-region
2
+
3
+ Category: Infrastructure · Level: advanced · Time: 45 min
4
+
5
+ An online shop's accounts and orders service runs in a single region today. Design how it will serve customers from three regions, keep working when one whole region is lost, and get there from where it is now without taking the shop down.
6
+
7
+ Before you design, find out where customers are and how much traffic each region gets, which data must be strictly correct and which can briefly disagree between regions, how much data could be lost and for how long the service could be down if a region fails, what the current system looks like, and what laws say about where data may live.
8
+
9
+ ## If you ask
10
+
11
+ - Today: one region in the US, one relational database of 20 TB with a standby copy, 30,000 requests per second at peak. 40% of customers are in Europe and 20% in Asia.
12
+ - Target: under 100 ms per request for customers in each region, and if a region is lost, under 5 minutes of downtime and no lost orders.
13
+ - An order must never be placed twice or lost. Profile details such as a saved address may take a few seconds to show up elsewhere.
14
+ - European customers' personal data must be stored in Europe.
15
+ - The move must not stop the shop taking orders, and every step must be reversible. Out of scope: the product catalogue and search.
16
+
17
+ ## Follow-ups
18
+
19
+ - A customer in Europe and their partner in the US change the same saved address within one second. What does each see afterwards?
20
+ - Halfway through the move, the new region shows a small number of orders that don't match the old one. What do you do?
@@ -0,0 +1,20 @@
1
+ # D20 · Checkout at 10x load
2
+
3
+ Category: Infrastructure · Level: advanced · Time: 45 min
4
+
5
+ An online shop's checkout works today: web servers behind a load balancer, one main relational database with two read copies, a shared cache for product and price data, and a call to an outside payment provider for each order. A sale has brought ten times the usual traffic, and checkout is failing. Decide what to change today to get through the sale, and what to change this quarter so it doesn't happen again.
6
+
7
+ Before you design, find out the usual and current load, what the dashboards show for each part, which failures customers are seeing, what the payment provider allows, how long the sale lasts, and which parts of checkout matter most to the business.
8
+
9
+ ## If you ask
10
+
11
+ - Usually 2,000 checkouts per minute; now 20,000, for another 6 hours. Product page views are up 15x.
12
+ - The main database is at 95% CPU and its connections are used up; query times have gone from 5 ms to 2 seconds. The read copies are 40 seconds behind.
13
+ - The cache hit rate has dropped from 98% to 70% since prices were changed for the sale. Web servers are at 40% CPU but many requests time out after 30 seconds.
14
+ - The payment provider allows 500 calls per second and takes about 1 second per call. Some customers are charged but see an error page, and retry.
15
+ - No order may be lost or charged twice. The team can deploy small changes within an hour; nothing can be rebuilt today.
16
+
17
+ ## Follow-ups
18
+
19
+ - The sale ends. Name the measurements you would alert on so the team hears about this before customers do next time.
20
+ - Next year's sale is expected to bring 50x. Which of today's changes still hold, and which don't?
@@ -0,0 +1,20 @@
1
+ # D21 · Webhook delivery
2
+
3
+ Category: Infrastructure · Level: core · Time: 45 min
4
+
5
+ A software company lets its customers register a web address, and sends an HTTP request there whenever something happens in their account (an invoice is paid, a user signs up). Design the service that delivers these events. Customers' servers are outside your control: some are slow, some are down for days, and some return errors.
6
+
7
+ Before you design, find out how many events and customers there are, how quickly events must be delivered, whether order matters, what happens when a customer's server is down, how customers know an event really came from you, and what they can see about past deliveries.
8
+
9
+ ## If you ask
10
+
11
+ - 200,000 customers with 500,000 registered addresses; about 1 billion events a day, peaking at 50,000 per second.
12
+ - Most events should arrive within a few seconds. A customer whose server is down must still get every event once it recovers, for up to 3 days.
13
+ - Customers expect events about the same object (one invoice) to arrive in the order they happened.
14
+ - Customers can see each delivery attempt and its result, and can ask for events from a time range to be sent again.
15
+ - Out of scope: the systems that create the events, and customers' servers.
16
+
17
+ ## Follow-ups
18
+
19
+ - One large customer's server starts taking 30 seconds to respond to every request. What happens to everyone else's deliveries?
20
+ - A customer accidentally registers an address that points at one of your own internal services. What could go wrong, and what changes?
@@ -0,0 +1,20 @@
1
+ # D22 · Game leaderboard
2
+
3
+ Category: Social · Level: core · Time: 45 min
4
+
5
+ Design the leaderboard for a mobile game. Players earn points after each match and can see the top players, their own rank and their friends' ranks. A new board starts every Monday.
6
+
7
+ Before you design, find out how many players there are and how many matches end per second, how quickly a new score must show in the rankings, which views players open most, how ties are broken, and whether past weeks are kept.
8
+
9
+ ## If you ask
10
+
11
+ - 50 million weekly players and up to 30,000 matches finishing per second at peak, each adding points to one player's score.
12
+ - Players see the top 100, their own exact rank (for example 4,381,207th), and the ranks of up to 500 friends.
13
+ - A new score should appear in the rankings within a few seconds. Ties go to whoever reached the score first.
14
+ - The board resets every Monday at 00:00 UTC, and each past week's final top 1,000 is kept for a year.
15
+ - Out of scope: matchmaking, detecting cheating, and rewards.
16
+
17
+ ## Follow-ups
18
+
19
+ - The game launches separate boards for each of 200 countries and each of 10 skill levels. What changes?
20
+ - Cheaters are found after the week has ended, and their scores must be removed. What happens to everyone else's final ranks?
@@ -0,0 +1,21 @@
1
+ # D23 · Discovery recommendations
2
+
3
+ Category: ML · Level: advanced · Time: 45 min
4
+
5
+ A shopping app adds a "For you" shelf on its home screen, showing items each person is likely to want. People don't follow anyone, so the shelf must pick from the whole catalogue using what each person has viewed, searched and bought, including what they did a few minutes ago.
6
+
7
+ Before you design, find out how many people and items there are, how many items the shelf shows, how quickly it must load, how soon a new action should change it, how often the catalogue changes, and how success is measured.
8
+
9
+ ## If you ask
10
+
11
+ - 80 million daily users and 50 million items for sale; about 2 million items are added or sold out each day.
12
+ - The shelf shows 20 items and must load within 200 ms. It is opened about 400 million times a day.
13
+ - Something a person views or buys should change their shelf within a minute.
14
+ - A data science team supplies models that score how likely a person is to buy an item; they are too slow to score more than a few thousand items per request.
15
+ - Success is measured as purchases from the shelf. Out of scope: how the models are trained, and ads.
16
+
17
+ ## Follow-ups
18
+
19
+ - A new user opens the app for the first time. What do they see, and why?
20
+ - Shoppers complain the shelf keeps showing items they have already bought. What changes?
21
+ - The team wants to compare two versions of the shelf on live traffic. What does the system need?
@@ -0,0 +1,20 @@
1
+ # D24 · Splitting a shared database
2
+
3
+ Category: Data · Level: advanced · Time: 45 min
4
+
5
+ Two teams at an online shop, orders and catalogue, share one large relational database, and they keep slowing each other down. Plan and design moving all order data into a store of its own while the shop keeps taking orders, with a way back at every step if something goes wrong.
6
+
7
+ Before you design, find out how much order data there is and how fast it changes, which other code reads or writes the order tables, which queries join order and catalogue data, how much downtime is acceptable, and how you would know the two copies agree.
8
+
9
+ ## If you ask
10
+
11
+ - 15 TB of order data in 40 tables, with 5,000 writes per second at peak. The catalogue tables are 3 TB.
12
+ - About 30 services read the order tables directly, and several reports join orders with products in a single query.
13
+ - The shop may never stop taking orders, and no order may be lost or duplicated. A few seconds of slower responses are acceptable.
14
+ - The orders team owns the work and has six months. Every step must be reversible until the very last one.
15
+ - Out of scope: redesigning how orders work, and choosing a vendor for the new store.
16
+
17
+ ## Follow-ups
18
+
19
+ - Two weeks after the switch, finance finds that some refunds recorded during the move are missing from the new store. What happens now?
20
+ - Halfway through, the catalogue team asks to split their tables out the same way. What changes?
@@ -15,6 +15,21 @@ gap and the question they should ask themselves next time.
15
15
 
16
16
  Do not run commands and do not write files: print the review only.
17
17
 
18
+ ## The whiteboard
19
+
20
+ In the hub, an attempt can come with a whiteboard: components with their settings (replicas, sizes,
21
+ cache expiry, replicas and shards), arrows along the request path, group boxes and notes, given to you
22
+ as text with an id in [brackets] for each item. Pen drawings can't be read as text and are only counted.
23
+
24
+ - Read the board as part of the design, mainly for **high-level** and **deep-dive**: does every
25
+ requirement have a path on it, and does it hold at the person's own estimated load?
26
+ - Say where the board and the notes disagree (a component in one and not the other, a setting that
27
+ contradicts an estimate).
28
+ - Quote their own settings and labels, as you would quote their notes. Still never name the
29
+ component, setting or technique that would fix a gap.
30
+ - Attempts without a board are reviewed on the notes alone, as before. Never mark one down for having
31
+ no board.
32
+
18
33
  ## Scores
19
34
 
20
35
  Score each dimension 1 to 4, with a one-line reason that points at their own notes (quote a few
@@ -81,6 +96,13 @@ Gaps: <tags, or none>
81
96
  Next time: <one concrete habit for the next run>
82
97
  ```
83
98
 
99
+ When the attempt has a whiteboard, add one more header line right under `Next time`, so the hub can
100
+ outline what each reason is about:
101
+
102
+ ```
103
+ Board: <part> <ids>; <part> <ids> (for example: deep-dive n3 n5; trade-offs n5), or none
104
+ ```
105
+
84
106
  Then, briefly:
85
107
 
86
108
  1. **Scores, one line each**: the dimension, the score, the reason from their notes.
@@ -92,3 +114,51 @@ Then, briefly:
92
114
  5. **Did last time's feedback stick?** Compare with the earlier reviews of this prompt: say plainly
93
115
  whether their gaps came back or were fixed. Skip if this is the first run.
94
116
  6. **Next time**: one habit, the same as the header line, with a sentence on how to practise it.
117
+
118
+ ## Teaching the review
119
+
120
+ The hub asks for a teaching review: it sends the person's save points (drafts), how far they went in
121
+ the Help tab, and their finish answers. Teach, don't only state findings. Use these exact `##`
122
+ headings, so the hub can lay the review out.
123
+
124
+ **Open** with one or two plain sentences on how it went: the approach, and what cost the time.
125
+
126
+ **`## Between your drafts`**, right after the opening, only when there are two or more save points.
127
+ Name the drafts where something changed (for example `### Drafts 2 to 3: the estimates reached
128
+ storage`), quote only the changed lines (at most 12 lines in one block), and add **What it means:** in
129
+ one to three sentences on what the change shows about how they worked. Leave the section out with one
130
+ draft or none.
131
+
132
+ **Plan against what they finished.** In point 2 (Time), compare what they set out to cover with what
133
+ the notes and board ended up with. When they answered "What changed from your plan", use their words.
134
+
135
+ **`## What to work on, with examples`**, after point 4. One `###` entry per finding that has a gap tag
136
+ (from the scores, the time or the board), written out in full:
137
+
138
+ - **What we saw:** their words or board labels, quoted, or what the drafts showed.
139
+ - **Why it matters:** one to three sentences: what goes wrong, and what an interviewer makes of it.
140
+ - **Before** and **After**: the same short fragment (at most 6 lines each) on a different, made-up
141
+ system, such as a bird-sighting log for a hiking club: a line or two of requirements, numbers or
142
+ boxes, first done the weak way, then the better way.
143
+ - **How to practise:** a short drill in words, or another prompt from the bank by its id.
144
+ - **What to say:** one sentence they could say in the interview.
145
+
146
+ The fragment teaches the habit (numbers carried into the design, one part pushed to its limit, a cost
147
+ named), on a system that is not this prompt. It never shows this prompt's design, and never shows the
148
+ component, setting or technique that would fix this prompt's gap, so the rule at the top still holds.
149
+
150
+ **What I learned.** If they wrote one, say in one line whether it names the main gap.
151
+
152
+ At the end, after point 6, three short sections the hub shows beside the review, in this order:
153
+
154
+ 1. **`## How you solved it`**: a list with Result (*solved alone*; *solved with questions* when help
155
+ stopped at step 4 or earlier; *solved with help* from step 5 on), Help (the furthest step and its
156
+ name, or none), Moved you (only when help was used: the step that got them moving, from their
157
+ finish answer or the conversation), Time and Drafts. Add one line on why it counts as that result.
158
+ 2. **`## Gaps noted, and why`**: one line per tag in the Gaps header line: why this tag, in plain
159
+ words, and what "fixed" looks like. Then `Not tagged:` with a tag you considered and left out, and
160
+ why. Leave this section out when Gaps is none.
161
+ 3. **`## When it comes back`**: `- Redo:` a weekday and date, two or three days out (three when solved
162
+ with help; for solid, a look back in about a week), and how: from a blank page, no help, the target
163
+ time. `- Then:` a sibling prompt from the bank, by its id. `- Daily:` a two or three minute drill
164
+ for the main gap.
@@ -0,0 +1,29 @@
1
+ # P21 · Reply drafts for support agents
2
+
3
+ Category: AI product · Level: core · Time: 35 min
4
+
5
+ You are the product manager for the support tools at a home-repair booking service. Support agents answer about 4,000 customer messages a day: late technicians, rescheduling, refunds, prices. The head of support wants the tool to draft a reply for each message so agents answer faster. Decide whether this needs AI, design it, and say how you would know it is good enough to launch.
6
+
7
+ Before you answer, find out what "faster" is worth to the business, which messages agents handle today and how, what a wrong reply would cost, and what the agents themselves think of the idea.
8
+
9
+ ## If you ask
10
+
11
+ - The goal is to cut the time to first reply, now 6 hours on average, without hurting customer satisfaction.
12
+ - About half the messages are one of ten common questions; the rest vary a lot.
13
+ - Agents can already pick from 40 saved reply templates, but most type from scratch.
14
+ - A reply that promises a refund the policy does not allow costs the company money and trust.
15
+ - There is a budget for model calls, but no one has said how much each reply may cost.
16
+ - Out of scope: replacing agents or answering customers with no person in the loop.
17
+
18
+ ## What a strong answer covers
19
+
20
+ - Decides between a model, rules or better templates from the problem, and says why.
21
+ - Keeps a person in control where a mistake is costly, and says where that is in the flow.
22
+ - Says how drafts would be tested before launch, with a set of cases and a bar to clear.
23
+ - Names what happens when a draft is wrong or slow, and the fallback.
24
+ - Weighs the cost and latency of each draft against the time it saves.
25
+
26
+ ## Follow-ups
27
+
28
+ - Agents start sending drafts without reading them. What do you change?
29
+ - The model is twice as expensive as planned. What do you cut first?
@@ -0,0 +1,29 @@
1
+ # P22 · Sorting receipts by category
2
+
3
+ Category: AI product · Level: core · Time: 35 min
4
+
5
+ You are the product manager of an expense app for small businesses. Owners photograph receipts and then pick a category (travel, meals, supplies and so on) so their accountant can file taxes. Picking categories is the step people skip most. The team wants to sort receipts automatically "with AI". Decide whether that is the right tool, and design the feature.
6
+
7
+ Before you answer, find out who fixes a wrong category and when, how many categories there are, what data the app already has about each purchase, and what a wrong category costs the owner.
8
+
9
+ ## If you ask
10
+
11
+ - There are 14 categories; the accountant can rename them per business.
12
+ - Most receipts come with a card transaction that already carries the merchant name and type.
13
+ - Owners sort about 60% of receipts today; the rest reach the accountant unsorted.
14
+ - A wrong category is usually caught by the accountant at the end of the quarter, which costs them time.
15
+ - Reading the photo itself is already done by a text-recognition service the app pays for.
16
+ - Out of scope: changing how receipts are captured.
17
+
18
+ ## What a strong answer covers
19
+
20
+ - Checks what simpler options (rules on the data already there, the owner's past choices) would get before reaching for a model.
21
+ - Designs what the owner sees: when a category is suggested, how sure it looks, and how to fix it.
22
+ - Says how it would be tested before launch and what share right is good enough, with a reason.
23
+ - Says what happens when the category is wrong, and who catches it.
24
+ - Accounts for cost and speed per receipt at the app's volume.
25
+
26
+ ## Follow-ups
27
+
28
+ - Accountants want to approve every automatic category. Does the design still help?
29
+ - A new kind of business joins with categories you have never seen. What breaks?
@@ -0,0 +1,31 @@
1
+ # P23 · A symptom helper in a pharmacy app
2
+
3
+ Category: AI product · Level: advanced · Time: 35 min
4
+
5
+ You are a product manager at a pharmacy chain whose app is used to refill prescriptions and buy over-the-counter medicine. Leadership wants a chat helper where people describe a symptom and get suggestions for what to buy, or a prompt to see a pharmacist or a doctor. A competitor launched something similar last month. Decide whether and how to build it.
6
+
7
+ Before you answer, find out what leadership hopes to gain, who would use it, what the pharmacists think, what the law allows the app to say, and what happens today when someone is unsure what to buy.
8
+
9
+ ## If you ask
10
+
11
+ - The hoped-for gain is more over-the-counter sales in the app, now 8% of the chain's sales of those products.
12
+ - Pharmacists already answer questions in store and by phone, and they are often busy.
13
+ - Legal says the app may suggest products but must never read as a diagnosis, and must send urgent cases to a person.
14
+ - Some users are buying for children or older relatives.
15
+ - The competitor's helper was criticised in the press for a wrong suggestion in its first week.
16
+ - Out of scope: prescription medicine and booking doctor visits.
17
+
18
+ ## What a strong answer covers
19
+
20
+ - Questions whether a model is needed for each part of the flow, and keeps the risky parts out of its hands.
21
+ - Treats a wrong answer as the main design problem: who is hurt, how it shows up, and the way out to a person.
22
+ - Sets tests before launch with cases chosen for their risk, and a launch bar stricter than average accuracy.
23
+ - Names online measures after launch, with guardrails for safety as well as sales.
24
+ - Weighs speed, cost and the press risk against the gain, and says what it would not launch.
25
+ - Makes a clear call, including "not yet" if that is where the reasoning leads.
26
+
27
+ ## Follow-ups
28
+
29
+ - The tests pass, but pharmacists say the suggestions are bland and unhelpful. What now?
30
+ - A user types a symptom that could be an emergency at 2 a.m. Walk through what happens.
31
+ - Leadership wants it live in four weeks to match the competitor. What do you say?
@@ -0,0 +1,29 @@
1
+ # P24 · Writing listing descriptions for hosts
2
+
3
+ Category: AI product · Level: advanced · Time: 35 min
4
+
5
+ You lead the host tools at a holiday-rental marketplace with about 200,000 listings. Many listings have short or out-of-date descriptions, and guests who book them complain more often. The team proposes that a model write each description from the host's photos, amenities and past reviews, which the host can then edit. Decide whether to build this and how.
6
+
7
+ Before you answer, find out how descriptions affect bookings and complaints, what data each listing has, what hosts would accept, and who is responsible when a description is wrong.
8
+
9
+ ## If you ask
10
+
11
+ - Listings with descriptions under 50 words get 15% fewer bookings and more complaints about "not as described".
12
+ - Amenities are ticked boxes the host fills in; about a fifth are out of date.
13
+ - Hosts in some countries must not make claims about a home they cannot back up.
14
+ - Writing a description takes the model a few seconds; regenerating all listings each month would cost a noticeable share of the host-tools budget.
15
+ - Guests cannot tell who wrote a description today.
16
+ - Out of scope: translating descriptions.
17
+
18
+ ## What a strong answer covers
19
+
20
+ - Separates what the model may say from what must come from the host, and says why.
21
+ - Designs the host's side: what they see, what they must confirm, and what happens if they ignore it.
22
+ - Tests before launch for the mistake that matters most (claims that are not true), not only for writing quality.
23
+ - Names how success and harm would be measured after launch, for hosts and guests.
24
+ - Reasons about cost and when to generate (once, on change, on a schedule) against the value.
25
+
26
+ ## Follow-ups
27
+
28
+ - A guest books because the description said "sea view", and there is none. What does the product do?
29
+ - Half the hosts accept the draft without reading it. Is that a success?
@@ -0,0 +1,28 @@
1
+ # P25 · A requirements doc for group bookings
2
+
3
+ Category: Written doc · Level: core · Time: 60 min
4
+
5
+ Take-home exercise. You are a product manager at a fitness-class booking app. Today each person books their own spot. Studios say friends often want to book together and give up when the class fills before everyone has booked. Write a product requirements document for group bookings, for the engineering lead and the head of studio partnerships, who will decide from it whether to build this next quarter. Keep it to about two pages.
6
+
7
+ Before you answer, find out how often friends book together today, what studios want and fear, what the app already supports, and what "next quarter" can hold.
8
+
9
+ ## If you ask
10
+
11
+ - About 18% of bookings are made within two minutes of another booking for the same class, often from the same card.
12
+ - Studios worry that one person will hold several spots and then not show up.
13
+ - The app already has waitlists, and a cancellation fee that each studio sets.
14
+ - One squad of five engineers is free next quarter.
15
+ - Out of scope: private classes and corporate accounts.
16
+
17
+ ## What a strong answer covers
18
+
19
+ - States the problem with evidence and who has it before proposing anything.
20
+ - Makes the proposal specific enough that an engineer could estimate it and a studio could object to it.
21
+ - Uses non-goals to mark the edges of the first version.
22
+ - Says what is being decided and what the author recommends, so a reader who was not there could decide.
23
+ - Keeps to the length, in full sentences a reader can follow without the author.
24
+
25
+ ## Follow-ups
26
+
27
+ - The engineering lead says it is twice the size you thought. What do you cut, and where in the doc does that show?
28
+ - A large studio chain says it will opt out. Does the doc change?