Hard examples — put Alpha to the test

Easy demos prove nothing. Here are 15 genuinely tough tasks — the kind students, developers and small-business owners in India actually bring: proofs, multi-file bugs, physics, long PDFs, contracts, live research and full app builds. Tap one and Chat or Build opens with the prompt ready. You read it, tweak it, and press send.

Honest note: we don’t publish benchmark scores, and these aren’t cherry-picked screenshots. Answers can vary between runs and modes, and AI can be wrong — so every card tells you how to check the result yourself. For the hardest ones, pick Forge in the mode picker.

Complex math

Proofs and multi-step derivations where a single slip shows. Every one has an answer you can check.

Olympiad number theory

Chat

Needs a real proof, then a counting step on top of it.

See the prompt
Prove rigorously that n^5 − n is divisible by 30 for every integer n.
Use modular arithmetic for 2, 3 and 5 separately (cite Fermat's little theorem where you use it) and explain why that is enough.

Then answer: how many integers n with 1 ≤ n ≤ 1000 make n² + 1 divisible by 5? Show the residue table mod 5 and the count.

Finally, write a 5-line Python check for both results.

How to check: Second answer is 400 (n ≡ 2 or 3 mod 5). Run the Python it writes.

Expected value with states

Chat

Easy to get wrong without setting up a proper Markov chain.

See the prompt
A fair six-sided die is rolled until two sixes appear in a row.
1) Set up the states and the system of equations for the expected number of rolls, and solve it exactly.
2) Generalise to k consecutive sixes and give a closed form.
3) Sanity-check with a short Monte-Carlo simulation in Python (state the expected spread for 100,000 trials).
Explain each step as if teaching a JEE aspirant.

How to check: Two sixes in a row → exactly 42 rolls; general form (6^(k+1) − 6) / 5.

A classic hard integral

Chat

Needs a symmetry trick, not a lookup — and a numeric cross-check.

See the prompt
Evaluate I = ∫₀^{π/2} ln(sin x) dx exactly.
Use the substitution x → π/2 − x and the double-angle identity; justify why the improper integral converges.
Then compute I numerically to 6 decimal places with a method of your choice, and confirm it matches your exact answer.
Finally, use the result to evaluate ∫₀^{π} x·ln(sin x) dx.

How to check: I = −(π/2)·ln 2 ≈ −1.088793; the last one is −(π²/2)·ln 2.

Large-codebase debugging

Multi-file bugs with several interacting causes — or your own repo, pasted file by file.

UPI checkout: 3 bugs, 4 files

Chat

Float money, a shared cache key and an async race — all at once.

See the prompt
You are a senior engineer debugging a production Node.js + TypeScript checkout service for an Indian D2C store. Below is a 4-file excerpt.

Customers report:
1) Some UPI totals are off by ₹0.01.
2) Occasionally a coupon is applied twice.
3) Rarely, a shopper sees another shopper's cart.

// money.ts
export const addGst = (rupees: number) => rupees * 1.18;
export const round2 = (n: number) => Math.round(n * 100) / 100;

// cart-cache.ts
const cache = new Map<string, Cart>();
export function getCart(req: Req) {
  const key = req.headers["x-forwarded-for"] ?? "anon";
  if (!cache.has(key)) cache.set(key, { items: [], coupon: null });
  return cache.get(key)!;
}

// coupon.ts
export async function applyCoupon(cart: Cart, code: string) {
  if (cart.coupon) return cart;
  const c = await db.coupons.find(code);   // ~40 ms
  cart.coupon = c;
  cart.discount = (cart.discount ?? 0) + c.amount;
  return cart;
}

// checkout.ts
export async function total(cart: Cart) {
  let sum = 0;
  for (const it of cart.items) sum += round2(addGst(it.price * it.qty));
  return round2(sum - (cart.discount ?? 0));
}

Tasks:
- Find the root cause of EACH symptom; point to exact lines and explain the mechanism (not just "race condition").
- Rank by severity (money / privacy / UX).
- Give minimal diffs that fix them (integer paise, a proper session key, an idempotent coupon apply).
- Write 3 focused Vitest tests that fail before your fix and pass after.
- List anything you would still check in the rest of the codebase that this excerpt cannot show.

How to check: A good answer finds: float rupees vs paise, cart keyed by IP (shared NAT/CGNAT), and a check-then-await race in applyCoupon.

Your own codebase

Chat

Map the architecture first, then trace the bug with evidence.

See the prompt
I'll paste (or attach) several files from my project. Act like a careful staff engineer joining the codebase.

Step 1 — Map: list each file's responsibility, the main data flow, and where state lives. Do not guess about files I haven't shared; ask for them.
Step 2 — Bug hunt: the bug is: [describe what you see, expected vs actual, steps to reproduce, any error/stack trace].
Trace the request through the files, form 2–3 hypotheses, and say what evidence in the code supports or rules out each.
Step 3 — Fix: give the smallest diff for the most likely cause, plus a test that would have caught it.
Step 4 — Risks: other places in the shared code with the same pattern.

Files:
[paste files here, each starting with its path]

You bring: Paste or attach your files (code, logs, stack traces).

How to check: Insist it cites file + line for each claim and asks for files it hasn't seen.

Build an app with its own tests

Build

Real algorithm + edge cases + a visible test panel in the preview.

See the prompt
Build a "Goa trip expense splitter" web app.
- Add friends, add expenses (who paid, amount in ₹, split equally or by custom shares).
- Compute balances and then MINIMISE the number of settle-up payments (greedy max-creditor / max-debtor), showing each step.
- Store everything in localStorage; Indian number formatting (₹1,25,000).
- Include a visible "Run tests" panel with at least 8 tests: rounding to paise, custom shares that don't add up, one person paying everything, zero-balance friends, and that total paid always equals total owed.
- Clean, mobile-first UI.

How to check: Click Run tests in the preview; then add a tricky expense and re-run.

Scientific reasoning

Derivations, estimates and experiment design — where assumptions matter as much as the answer.

Why the sky is blue (properly)

Chat

Dimensional analysis → a number → a real-world explanation.

See the prompt
Explain why the daytime sky is blue but sunsets over the Arabian Sea look orange-red.
1) Derive the λ⁻⁴ dependence of Rayleigh scattering using dimensional analysis (state what the scattered intensity can depend on).
2) Compute how many times more strongly 450 nm light scatters than 700 nm light.
3) Explain why the sky isn't violet even though violet scatters even more.
4) Explain why big particles (dust, pollution haze in Delhi winter) make the sky look whiter instead.
State every assumption.

How to check: (700/450)⁴ ≈ 5.9×. Violet: weaker sunlight + eye sensitivity.

Fermi estimate: India's chai energy

Chat

Chaining assumptions with units — and bounding the error.

See the prompt
Estimate the energy used to heat water/milk for all the chai India drinks in one day.
- List every assumption (people, cups per day, ml per cup, temperature rise, stove efficiency) with a low / likely / high value.
- Show the calculation with units at each step, in kWh.
- Compare the likely value to the daily output of a 1 GW power plant.
- Say which single assumption moves the answer most, and by how much.

How to check: Check the units line by line; specific heat of water ≈ 4.18 kJ/kg·°C.

Design a real experiment

Chat

Controls, confounders, statistics — not just a hypothesis.

See the prompt
Design a school/college-lab experiment to test whether turmeric (curcumin) inhibits growth of E. coli on agar.
Include: hypothesis and null hypothesis, independent/dependent/controlled variables, positive and negative controls, how to handle curcumin's poor water solubility (and a solvent-only control), replicates and a rough sample-size argument, the statistical test you'd use and why, safety notes, and what result would FALSIFY the hypothesis.
End with 3 ways the experiment could give a misleading result.

How to check: Look for a solvent-only control and a stated statistical test.

Long-document analysis

Bring a long PDF, contract or web page. Ask for quotes and page references so every claim is checkable.

Annual report / paper → brief

Chat

Find numbers, contradictions and gaps across many pages.

See the prompt
I've attached a long document. Produce a structured brief:
1) 10-line executive summary.
2) Table of the 10 most important numbers: value | what it measures | page/section.
3) Key risks or limitations the document itself mentions.
4) Any statements in different sections that contradict or don't quite match each other.
5) 5 important questions this document does NOT answer.
Rules: quote the exact sentence for every factual claim, cite the page/section, and write "not in document" instead of guessing.

You bring: Attach a PDF or Word file with the + button first.

How to check: Spot-check 3 quotes against the PDF pages it cites.

Rental agreement / offer letter review

Chat

Obligations, deadlines and unusual clauses, in plain English.

See the prompt
I've attached an agreement (rental agreement / job offer / vendor contract) from India. Read all of it, then give me:
- Each party's obligations, as two lists.
- Every deadline, notice period, lock-in, deposit and penalty, with clause numbers.
- Clauses that look unusual or one-sided compared with common Indian practice — explain why, quoting the clause.
- Things that are missing that I would normally expect.
- 8 questions I should ask before signing.
This is for understanding only, not legal advice — flag anything where I should consult a lawyer.

You bring: Attach the agreement (PDF / Word / photo) first.

How to check: Every point should carry a clause number you can find in the file.

Read a long web page

Agent

Uses page reading on a long article and builds a timeline from it.

See the prompt
Read https://en.wikipedia.org/wiki/Chandrayaan-3 in full.
Then give me:
1) A dated timeline of the mission from launch to the end of surface operations.
2) A table of the mission's payloads: name | on lander or rover | what it measured | key finding (if reported).
3) What the page says went differently from Chandrayaan-2, and why that mattered.
Only use what the page says, and mark anything you add from general knowledge.

How to check: Compare 3 timeline dates with the article.

Multi-step agent tasks

Agent mode plans, searches, reads pages and assembles the result step by step. Build turns a spec into a working app.

Research & compare payment gateways

Agent

Several searches, page reads, a table and a recommendation.

See the prompt
I run a small D2C store in India doing about ₹2 lakh GMV per month (mostly UPI, some cards).
Research the current published pricing and terms of Razorpay, Cashfree Payments and PhonePe Payment Gateway:
- standard fee for UPI, domestic cards and international cards, settlement time, setup/annual fees, and anything notable for small merchants.
Search the web, read each provider's official pricing page, and cite the source URL for every number.
Then build a comparison table, estimate my monthly fees on each assuming 80% UPI / 20% domestic cards, and recommend one with reasons.
End with a list of things I must confirm with the provider, because pricing changes.

How to check: Open the cited pricing pages; numbers must match today's page.

Plan a budget trip with live info

Agent

Plan → search → constrain → re-plan within a budget.

See the prompt
Plan a 4-day trip from Bengaluru to Hampi for 3 college students, total budget ₹12,000 for all three (travel, stay, food, entry tickets).
Steps: search current train/bus options and approximate fares, find 2–3 budget stays with indicative prices, list the must-see sites with entry fees and opening times, and check the weather for this season.
Then produce a day-by-day plan with a cost table that stays under budget, and a plan B if it rains. Cite sources for prices and mark anything that's an estimate.

How to check: Check the totals add up and the sources are real pages.

Spec → working GST invoice app

Build

Domain rules (CGST/SGST vs IGST), number-to-words in lakh/crore, print layout.

See the prompt
Build a GST invoice generator for a small kirana / retail shop in India.
- Shop profile (name, GSTIN, state) saved in localStorage.
- Line items: name, HSN code, qty, rate, GST rate (0/5/12/18/28%).
- If buyer state = shop state → split tax into CGST + SGST; otherwise IGST. Show tax per rate slab.
- Grand total in figures (₹1,23,456.00) and in words using Indian numbering (lakh/crore), e.g. "Rupees One Lakh Twenty-Three Thousand ...".
- Validate GSTIN format (15 chars) and show friendly errors.
- Invoice history, and a clean print-ready A4 layout (window.print).
- Mobile-first, works offline.

How to check: Try an inter-state and an intra-state invoice and check the tax split.

How these open

  • Chat cards open a fresh chat with the prompt in the box — nothing is sent until you press send.
  • Agent cards also switch on Agent mode, so Alpha can search the web and read pages step by step.
  • Build cards open a new Build with the spec ready; you get a live preview and can download the code.
  • Free demo limits apply; sign in with Google for more. See how Alpha works and pricing.