YC startup ideas: Y Combinator's current Requests for Startups and what a small builder can make from each

Build a web app with no login.
User: a support worker helping a client fill in one specific form.
Screen: a short interview, one question at a time, that ends in a list of what is missing or inconsistent.
Data: a markdown file of the official instructions for this form, which I will add. Every finding must quote the sentence of the instructions it comes from.
Do not build: accounts, document upload, advice outside the instructions.
Done when: for 5 sample cases I provide, the tool flags the missing items I expect and quotes the right sentence each time.

Build a mobile-friendly web app. Keep everything on the device; no server storage of personal documents.
User: someone renewing one benefit program in one state.
Screen: a checklist of required documents from the program's public instructions, a place to note where each file is, and a renewal date reminder as a downloadable calendar file.
Data: a JSON file of required documents and deadlines that I will fill in from the agency website.
Do not build: accounts, document upload, eligibility decisions.
Done when: I can tick off every item, reload and see my progress, and download one calendar entry.

Build a web app with no login.
User: a homeowner preparing a permit application for one project type in one city.
Screen: paste or upload the application text; get a table of checklist items marked found, not found or unclear, each with the quoted passage.
Data: the city's public checklist as a markdown file I will add.
Do not build: drawing analysis, code-compliance verdicts, accounts.
Done when: on 5 real sample applications, every item I know is missing is marked not found, and no item is marked found without a quote.

Build a web app.
User: a reader judging writing quality.
Screen: two anonymous passages answering the same prompt; buttons for Left, Right, Both bad. After a vote, reveal which model wrote which.
Data: a JSON file of prompts and pairs of outputs that I will provide. Randomize left and right.
Do not build: accounts, model calls, comments.
Done when: 20 votes are stored, a results page shows win rate per model, and the left-right split in the data is within 45 to 55 percent.

Build a small benchmark site.
User: a researcher comparing methods on one forecasting task.
Data: a public time-series dataset I will add, split into a visible training part and a hidden test part.
Screen: task description, a download link, a form to submit predictions as a CSV, and a leaderboard sorted by error.
Scoring: mean absolute error, computed on the server against the hidden part.
Do not build: accounts, model hosting.
Done when: submitting a file of the training mean as the prediction gives the baseline score shown on the board.

Build a command-line tool and a local web page.
User: an analyst who may not send documents outside the building.
Behavior: index a folder of text and PDF files; answer questions with a quoted source for every answer, using a locally run open-weight model.
Constraint: the app must work with all network access disabled. Add a startup check that fails loudly if any outbound connection is attempted.
Do not build: accounts, cloud sync, telemetry.
Done when: with the network off, 5 questions about 20 sample files return answers with quotes, and the outbound-connection check passes.

Build a web app for a small team.
User: an engineer comparing models on one internal task.
Screen: a table of test cases (input, expected output); a Run button that sends each input to a model I configure and marks pass or fail; a page that lists every failure.
Data: one SQLite file. API keys stay on the server.
Do not build: model training, accounts beyond a shared password.
Done when: I import 20 cases, run two models and see a pass rate for each plus the failing cases side by side.

Build a single-page calculator. No backend.
User: someone testing an orbital business idea.
Inputs: payload mass in kg, launch cost per kg, years of service, yearly revenue, yearly operating cost.
Output: total cost, total revenue, break-even launch cost per kg, and a table that shows the result at half and double each input.
Leave all inputs empty by default; show a prompt to enter your own numbers.
Done when: payload 100, cost per kg 5000, 5 years, revenue 200000 a year, operating cost 50000 a year shows break-even at 7500 per kg.

Build a single-page tool. No backend.
User: a hardware founder planning a robot.
Screen: an editable table (part, supplier, country of origin, unit cost, quantity). Show total cost, domestic cost, domestic percent, and a field for the target percent that I type myself.
Do not build: supplier databases, accounts, any built-in threshold.
Done when: three parts costing 100 each, one marked domestic, show 33.3 percent domestic, and the page says whether that is above or below the target I entered.

Build a single-page calculator. No backend.
User: a small business owner deciding between buying and leasing a machine.
Inputs: price, interest rate, loan years, lease payment per month, lease months, expected resale value, expected extra revenue per month.
Output: total cost of each option, the month in which each option pays back, and a plain-language note on which wins and why.
Add a visible note: general information, not financial advice.
Done when: price 60000, 8 percent over 5 years gives a monthly payment of about 1217.

Build a small web app.
User: a procurement engineer comparing delivery quotes.
Screen: a form (equipment type from a fixed list, quoted lead time in weeks, month of the quote, region) and a table with the median and range per type and quarter.
Rules: no company names, no free-text field, rate-limit submissions by IP, show a type's numbers only once it has at least 5 reports.
Do not build: accounts, comments.
Done when: 5 test reports for one type show a median, and 4 reports show 'not enough reports yet'.