LearnThe shelf

Prompt kit

  • 19 min
  • 13 stops

Each card is a prompt to paste into an AI assistant, with what a good answer contains so you can reject a bad one. An AI can be wrong; check its claims against the lesson.

Lesson 1: Level and tier

1 card.

Ask your AI · Level my evidence
Code, 16 lines
Act as a hiring-committee member levelling a software engineer in
India. Use four rows: scope (what I own), ambiguity (how vague the
problems I am handed are), influence (whose work changes because of
me) and execution (what I deliver, and with how much help).
My years of experience: [ ]. Current title and company type: [ ].
My last two years, one item per line, with numbers where I have
them: [projects, on-call, designs, mentoring, migrations]
1. Put each item in one row and say which level it reads like:
   SDE-1, SDE-2, Senior or Staff. Quote my words as the evidence.
2. Give the level most rows support, and its neighbour.
3. Translate both to an Indian product company, a global firm's
   India centre and an IT services firm, saying these mappings are
   approximate and not any company's official map.
4. Name the row with the weakest evidence and what would move it.
5. Flag any place where my years, not my evidence, did the work.
Do not quote salary figures.

A good answer contains: each item in a row with your words quoted; a level and a neighbour, not one verdict; cross-tier names hedged as approximate; the weakest row with a concrete action; and a line saying years are not evidence. Test it on Asha first (the years-against-scope table in lesson 1): a good answer reads SDE-2 with a case for Senior, influence the gap. If it calls her Senior for her years, or prints a pay band, do not trust it with yours.

Next stop: Lesson 2: Money, once, 2 min

Lesson 2: Money, once

1 card.

Ask your AI · Check my in-hand, both regimes
Code, 20 lines
Act as a careful payroll checker in India for tax year 2026-27
(Income-tax Act, 2025). Use only these rules unless I say otherwise.
New regime: nil to 4 lakh, then 5, 10, 15, 20, 25% in 4-lakh steps
to 24 lakh, 30% above; standard deduction 75,000; rebate up to
60,000 if taxable income is 12 lakh or less, marginal relief above;
no HRA, PF or professional-tax deduction.
Old regime (under 60): nil to 2.5 lakh, 5% to 5 lakh, 20% to 10
lakh, 30% above; standard deduction 50,000; deduct professional tax,
my PF (cap: check the current rule) and the deductions I list; rebate up to
12,500 at 5 lakh or less. HRA exemption, old regime only: least of
HRA, rent minus 10% of basic, and 50% of basic in Mumbai, Kolkata,
Delhi, Chennai, Hyderabad, Pune, Ahmedabad or Bengaluru, else 40%.
Cess 4% on tax in both regimes.
My facts, per year: basic [ ], HRA [ ], special allowance [ ],
other allowances [ ], employer PF [ ], gratuity provision [ ], my
PF [ ], professional tax [ ], rent paid [ ], city [ ], other
old-regime deductions with amounts [ ].
Show each step in a table, then my in-hand a year and a month under
each regime. List every assumption you made, and say if any rule
you used dates from before April 2026.

A good answer contains: gross as earnings only, with employer PF and the gratuity provision kept out of taxable salary; the new-regime steps slab by slab, with rebate or marginal relief and cess; all three HRA figures with the smallest picked; professional tax and your PF deducted in the old regime only; in-hand as gross minus your PF, professional tax and tax; and a list of assumptions. A weak answer uses four cities, gives HRA or PF deductions in the new regime, or offers one number with no steps.

Next stop: Lesson 3: Resume and calls, 2 min

Lesson 3: Resume and calls

3 cards.

Ask your AI · Interrogate my bullets
Code, 7 lines
Play a sceptical senior interviewer at a product company in
India. For each bullet below, ask me the first two follow-up
questions you would ask, then tell me which number or fact in
it I must be ready to defend and what evidence would satisfy
you. Do not rewrite the bullets.
Bullets:
[paste your bullets, one per line]

A good answer contains: questions of the kinds "measured how, over what window?", "what was your part?" and "what alternative did you reject?"; a flag on every number with no stated basis; and a flag on every "led" or "we" that hides who did what. Judge yourself, not the model: for each follow-up, do you have a true answer you could give in two sentences?

Ask your AI · Senior resume panel, one of three mock loops
Code, 7 lines
Play three people in turn: a recruiter who skims my top third
for 20 seconds, a hiring manager who reads every bullet, and a
referee-caller who rings my former manager. My target is an
SDE-3 role. For each, ask me three questions and say what you
would write about me. Focus on decisions, people beyond me, and
the number a year later. Do not rewrite my bullets.
My bullets: [paste]. My stints and gaps: [paste].

A good answer contains: nine questions, three notes in the three voices, a flag on every number with no stated basis, and one line for any bullet that shows a task but no decision. A partner can run the same card: read each voice aloud, time the recruiter to 20 seconds, and hand back the notes. As one of three mock loops, run this one, the recruiter-call card in lesson 4 and the bullet card above, in that order, on three different days.

Ask your AI · Will a recruiter's search find me?
Code, 12 lines
Act as a recruiter in India using a candidate database. Your
client brief: [stack keywords], [experience range], [city],
notice up to [N] days, salary [range] LPA.
My profile fields: headline [ ]; experience [ ]; city [ ];
notice period [ ]; current CTC [ ]; expected CTC [ ].
1. Would I appear in your search? Name each field that would
   remove me. If it depends on whether the salary filter reads
   current or expected CTC, say so; do not assume which.
2. Which words of my headline match your keywords, and which
   search words are missing from it?
3. If a field is true and still removes me, which other routes
   could reach this client?

A good answer contains: the field that filters you out, usually notice or salary; a plain statement that it cannot know whether the salary filter reads current or expected CTC; headline words matched and missing; referrals and direct messages as the routes round a true long notice; and no suggestion to change a field so it no longer matches your documents. If it tells you to shorten your notice in the field, discard the answer.

Next stop: Lesson 4: The recruiter call, 4 min

Lesson 4: The recruiter call

4 cards.

Ask your AI · Recruiter-call rehearsal
Code, 17 lines
You are a corporate recruiter at an Indian software company
screening me for the level I name below. Ask me the nine
questions and two probes below, one at a time, then the three
pushes, and after each answer write the one line that the
hiring manager would see on the form. Then tell me the one
answer that would make you drop me. Questions: background,
current CTC, expected pay, notice period, relocation and
office days, other processes, reason for change, level,
process and timeline. Probes: rate me 1 to 10 on my main
skill, and why this company. Pushes: "we go up to 30 per cent
on current", "we don't share bands", and "will you join if we
match?". Once, say your budget is below my ask.
My facts: [level I want, current CTC components, notice period
per letter, location, other processes, reason for change, my
ask, the job posting]. Keep my floor out of the facts. If I am
SDE-3 or above, once push me one level down and see whether I
ask for the higher loop.

A good answer contains: the questions in order, one form line per answer, and a flag on any answer that invents an offer, promises a date you cannot keep, gives a rating with no evidence, names a hike percentage, or puts a floor in writing. Check any figure it gives against your own payslip.

Ask your AI · Senior screen rehearsal card
Code, 11 lines
Play a recruiter at an Indian product company screening me for
an SDE-3 role. Run three loops. Loop 1: the nine questions in
order, one form line per answer. Loop 2: push me toward SDE-2
once, ask about my 14-month stint [or my gap or layoff], and
ask if I would stay for a counter-offer. Loop 3: say the band is
mostly stock and ask me to compare it with my current pay, then
say this opening's budget is below my ask.
After each loop name the one answer that would make you drop me
and the signal you wrote for the hiring manager.
My facts: [fixed, variable, RSUs granted/vested/to vest, notice
per letter, the stint or gap, the level I want and why].

A good answer contains: three loops, the senior pushes in the right loops, a form line each time, and a flag if I named a date I do not control or invented an offer. A partner can do the same job from this card: read the questions aloud, tick the six rubric lines on cheat sheet C1, and swap roles. Treat its numbers as unchecked until compared with your own papers.

Ask your AI · Hiring-manager screen rehearsal
Code, 13 lines
Play the hiring manager for the team I am interviewing with,
in a 45-minute screen. You have my resume, pasted below. Ask
me to walk you through [my main system] and stop me at three
minutes. Then ask one follow-up per bullet: why I chose the
design over the obvious alternative, how I de-risked the
change, what broke, what a number in the bullet means, what I
would change, and how I grew the people I name. Then ask how
hands-on I am, a disagreement with a peer, a time the brief
was unclear, and one push on ordering or duplicates in my
system. After each answer, write the one line you would put
in your notes and the follow-up a sceptical manager would ask
next. At the end, name the bullet I defended worst.
My resume: [paste, with phone and email removed].

A good answer contains: a cut-off at three minutes, one follow-up per bullet, a note line for each answer, and at least one question that sets two of your own bullets against each other (Meera's was dual writes against dual publishing). Hold your answers to Meera's bar in lesson 4: the disagreement and the unclear brief in two sentences each, ending in a measure or a trade, as in her people paragraph; the push on duplicates or ordering named where it lands and what catches it, as in her "What broke?" answer. If it accepts a number you cannot explain, it is being too kind; ask it to press harder.

Ask your AI · Staff or Principal search call
Code, 10 lines
Play a search-firm partner calling me about a [Staff or
Principal] seat at an Indian company you do not name at
first. Open with "Tell me about yourself in a minute." Then
answer my questions only as far as a real search partner
would. Once, ask for my resume before naming the client;
once, ask for my current manager as a reference; once, say
the panel is calibrated at Senior. After each of my turns,
write the line you would send the client.
My facts: [my scope, the cross-team result in numbers, the
people I grew, whether my search is confidential].

A good answer contains: a pitch at org scope with a number and who did the work; lesson 4's four search-call questions (client, approved headcount, reporting line, who decides) asked before any resume goes out; the resume held until the client is named; no current-manager reference before you resign; and the Senior calibration questioned, not accepted. Flag it if the AI names a real company or treats a seat as approved without saying who approved it.

Next stop: Lesson 5: Prep plan, 1 min

Lesson 5: Prep plan

1 card.

Ask your AI · Blank-page drill
Code, 5 lines
I am preparing for a [Java 21] coding round. Give me one problem
in the [pattern] family, without the solution, and wait for my
attempt. When I send my code and a spoken explanation, score it on:
the state I named, the transition, the complexity, and the one edge
case I missed. Do not show the solution until I ask.

A good answer contains: the problem alone, with the solution held back; a score on each of the four named items; and the one edge case it saw in your text. It does not invent a retention percentage.

Next stop: Lesson 6: The coding round, 2 min

Lesson 6: The coding round

3 cards.

Paste these into an AI assistant with your own facts filled in; each says what a good answer contains. Do not accept a model's claim about how often a problem is asked at any company: no such number is in this lesson's sources.

Ask your AI · Pattern spotter
Code, 8 lines
Here is a problem: [paste]. Do not solve it yet. List the cues in the
wording that point to a pattern from this list: two pointers, sliding
window, prefix sums, intervals, monotonic stack, BFS/DFS, topological
sort, union-find, binary search on the answer, heaps (including two-heap
median), backtracking, DP, tries, bit tricks, Dijkstra / 0-1 BFS, sweep
line, Fenwick or segment tree, quickselect. Name the one you would try
first, say what would make it fail, and say what changes if the input
is too big for memory or arrives as a stream.

A good answer contains: the exact words it relied on, quoted from your problem; a second candidate pattern; and a condition under which the first would break (for example a sliding window fails when the window rule is not monotone). Reject any answer that names a pattern with no quoted cue.

Ask your AI · Trace it by hand
Code, 4 lines
Here is my code: [paste]. Take the input [a small list you choose, four
elements, one repeated]. Write a table of every variable after each loop
iteration. Then say whether the output matches what the problem asks, and
name the single line that would fail if the input were empty.

A good answer contains: a table with one row per iteration using your own variable names, and a concrete empty-input failure, not a generic warning. Check the table yourself on paper, because a model can write a table that is wrong.

Ask your AI · The cost sentence
Code, 3 lines
My solution is [paste code]. Derive its time and space complexity from
the code, one loop at a time, in two sentences a reviewer could follow.
Then say what a brute-force version would cost and what it wastes.

A good answer contains: each loop accounted for (what it counts over and whether the inner pointer restarts), the brute-force cost, and the waste named (repeated scanning, repeated sorting). Reject a bare "O(n log n)" with no derivation.

Next stop: Lesson 7: Machine coding and low-level design, 1 min

Lesson 7: Machine coding and low-level design

3 cards.

Next stop: Lesson 7: machine coding, 1 min

Lesson 7: machine coding

Ask your AI · Which class changes why
Code, 4 lines
I am practising a 90-minute machine-coding round on a parking lot.
Here is my class list: [paste]. Ask me, one question at a time,
which requirement would change each class. Do not write code.
After my answers, tell me which class has two reasons to change.

A good answer contains: questions about pricing, vehicle sizes and concurrency; one class named by its reasons to change; no code.

Ask your AI · Check my leave method
Code, 3 lines
Here is my leave method: [paste]. Check whether it changes any
state before every check has passed. List each line that mutates
state and each check, in order.

A good answer contains: the checks and the mutations listed in order; any mutation that comes before a raise flagged; no rewrite of the method unless you ask.

Ask your AI · Add a pricing rule
Code, 3 lines
Give me a flat-rate, a weekend-rate and a monthly-pass pricing
class that all satisfy this interface: [paste]. Then tell me
whether any of them needs a change to ParkingLot.

A good answer contains: a flat-rate class that needs no edit to the lot. Against lesson 7's fee(vehicle, minutes), the finding that a weekend rate and a monthly pass do not fit, because the rule sees no entry time and no plate. The fix it names: fee takes the ticket, which carries the plate and the entry time, so park, its callers and every rule change (lesson 7, "Now add X").

Next stop: Lesson 8: System design, 1 min

Lesson 8: System design

3 cards.

Next stop: Lesson 8: system design, 1 min

Lesson 8: system design

Ask your AI · Check my estimate
Code, 5 lines
I am practising a 45-minute system design round at Senior level.
Prompt: design a URL shortener. Here are my assumptions and my
estimate: [paste]. Check my arithmetic, using 86,400 seconds in a
day. For each number, tell me whether my "so" follows from it.
Do not redesign the system.

A good answer contains: a recomputed rate for each of my numbers, any rounding error with its direction and size, a note on each line that has no consequence, and no new design.

Ask your AI · Failure drill: news feed
Code, 5 lines
Play a strict interviewer for a news feed design. Ask me one
question at a time about failure: the queue redelivers, the
timeline cache is lost, a celebrity posts, a region is down.
After each answer say whether I named the user-visible effect and
a mitigation with a cost. Do not give me the answer first.

A good answer contains: one question per turn, a verdict on whether the effect and the mitigation were both stated, and a pointer to the building block that covers any gap, without writing the answer for me.

Ask your AI · Find the double-spend
Code, 5 lines
I am designing a payments ledger. My design: [paste]. Find three
places where a retry, a timeout or a partial failure could move
money twice or lose it. For each, name the mechanism that prevents
it (idempotency key, transaction, outbox, reversal) and tell me
which assumption I have not stated.

A good answer contains: three concrete failure sequences in order, the prevention mechanism for each, the unstated assumption (for example that the bank supports an idempotency key, or the key lifetime versus my retry window), and no claim that any company does it this way.

Next stop: Lesson 9: The AI-allowed round, 1 min

Lesson 9: The AI-allowed round

1 card.

Ask your AI · The AI-assisted round
Code, 12 lines
Act as the interviewer in a 60-minute AI-assisted coding
round. Write a three-file Java 21 service, about 150
lines, with one failing test and one hidden bug that
compiles. Show only the code; under it, write the
bug's cause in ROT13. Do not solve it. Ask me first to
explain the failure. Then answer my requests as a typical
assistant would, including one plausible mistake, such as
a changed rounding rule. At minute 60, score me 0 to 2
on: hypothesis first, rule clarified, prompts a reader
can follow, a separating test, the diff reviewed,
ownership aloud, a close inside the time. Quote my words,
and say whether I caught your mistake and how.

A good answer contains: a wait for your diagnosis before any help; the cause written in ROT13 under the code; a subtle planted mistake, such as a changed rounding rule, not a syntax error; a 0 to 2 score on each of the seven rubric lines with your words quoted; and no invented company rubric.

Next stop: Lesson 10: The behavioural round, 2 min

Lesson 10: The behavioural round

2 cards.

Ask your AI · Story scorer
Code, 15 lines
You are an interviewer scoring a behavioural answer on four
levels: outstanding, solid, borderline, poor. Criteria:
1 headline and stakes in the first two sentences
2 "I" for decisions
3 one option rejected with its cost
4 a number with a baseline and how it was measured
5 one reflection I have acted on
6 whether the story holds up after two probes ("what exactly
  did you do?" and "how did you know?")
Score each criterion with one line of evidence quoted from my
text, then say what would move it up one level.
My role and level: [paste]
My story: [paste]

A good answer contains: a quote from your text for every score; no invented facts; one named change per criterion. Reject any answer that rates a number you have not measured as outstanding, and any figure about "how often this scores" that names no source and date.

Ask your AI · Mock interviewer (lesson 5 gate: Mock 1 and part of Mock 3)
Code, 19 lines
Act as a behavioural interviewer at my level. Ask me one
question at a time from this list and wait: [paste 5 questions:
led without authority, lost a disagreement, a failure that cost,
a half-finished project, why this company].
After each answer ask the next probe from this ladder, one only:
what exactly did you do / what did the others do / where did the
number come from / what did you reject / who disagreed / what
happened six months on / what would you change / what if they
had said no.
Never help or praise during the answer. At the end, score each
answer on the seven rows (headline, "I", option rejected,
number, reflection, two probes, scope at my target level) as outstanding, solid, borderline or
poor, quoting my words for each, and name the single row to
fix first. Flag any number I did not say how I measured.
My level: [Intern, SDE-1, SDE-2, SDE-3, Staff or Principal]
(Staff and Principal: add "a bet that failed" (lesson 12) to the list).
My stories: [paste]

A good answer contains: one probe at a time and no coaching until the end; a quote for every score; no invented facts about my work; the weakest row named per answer. Reject a scorecard that praises without a quote.

Next stop: Lesson 11: The hiring manager round, 2 min

Lesson 11: The hiring manager round

2 cards.

Ask your AI · Approach question
Code, 8 lines
Act as an engineering manager. Ask me one "how would you
approach it" question about a project like [my last project,
two lines, no confidential names]. Let me answer in seven parts,
one at a time: clarify, success measure and limits, the shape
with one rejected option, rollout with a rollback, risks and
what I would measure, team and cost, my question back. After
each part, say in one line what you would write in your notes.
Do not invent facts about my employer or your own company.

A good answer contains: one part per turn; a note line that quotes my words; a push-back when I name a technology before the goal; no invented company process.

Ask your AI · Project walk and probes
Code, 14 lines
Act as the hiring manager for a [level] seat on a team that
owns [one line, no confidential names]. I will walk you
through one project of mine in ten minutes. Then ask these
probes one at a time, waiting for each answer: why not the
option I rejected; how much of it was mine; what breaks at
ten times the load; someone I grew, then "how do you know?";
a launch date one third shorter than I need. Push back once
on a number I give. At the end, score me on seven boxes,
each seen or not, quoting my words: seat, level and gap
asked; two decisions with a rejected option and its cost;
credit exact; a before-and-after number; the ten-times
answer from today's load; the push-back conceded with a
test; the doubt asked for. Do not invent facts about my
employer.

A good answer contains: one probe per turn; a push-back on a real number I gave, not a generic one; a score per box with my quoted words as the evidence; "not seen" where I did not do it, with no partial credit.