Under the hood
Sharp Questions is a studio for writing job descriptions and planning interviews. This page explains exactly how it works, what we store, what we log, and what the panel actually is.
The studio
When you sign in, Sharp Questions opens a Postgres row for your job description and a transcript of your conversation. The conversation drives the document: you answer pointed questions, and a large language model — Anthropic's Claude — writes and revises the eleven sections as you talk. The application is a Next.js app running on Vercel; the database is hosted on Neon. Every word of your job description is generated by Claude on the basis of what you've told it, the canonical structure of a complete job description, and the running context of the draft.
The panel
The six advisors — Sullivan, McCord, Horowitz, Martin, Dunford and Sutherland — are real people, all of whom have written and spoken extensively in public about hiring, strategy and persuasion. None of them is affiliated with Sharp Questions. None of them has reviewed your job description. They are not employees, contractors or partners.
What the studio calls “advice” from these advisors is an inference: a model reading your draft and producing the kind of note each individual might leave, given what each has said publicly across their books, interviews, talks and articles. It is the same logical move as “in the style of” — useful as a critical lens, not a representation. Notes are signed because attribution clarifies the lens; they are not endorsements, quotations or paraphrases of anything the named person has said about your specific role.
If you'd rather have unsigned advice, turn the panel off in your settings.
What we store
Every job description you create, every message in the conversation that produced it, and every piece of generated advice is stored in your account in the database. You can delete a job description at any time, which removes the document, the conversation and the associated advice from the database. We do not sell or rent your data to anyone. We do not use it to train models.
What we log
Application logs are kept for thirty days to help us debug and to detect abuse. They contain HTTP request metadata (timestamp, route, response code, request duration, IP address), application errors, and the database row IDs touched in a request. They do not contain the contents of your job descriptions or conversations — those live only in the database.
Requests to Anthropic carry the prompt and the document context required to produce the next response. Anthropic's data handling is governed by Anthropic's terms; we don't send anything to them that you wouldn't expect us to send.
Cookies
We set three first-party cookies. A session cookie keeps you signed in. A CSRF token protects state-changing requests. A small preference cookie remembers your theme. That's all. No analytics tags, no third-party trackers, no advertising IDs.
Where it runs
The application runs on Vercel. The database runs on Neon. Anthropic processes model requests on its own infrastructure under its enterprise terms. The marketing page you're reading is a static export served from Vercel's edge.
What this adds up to
You're writing in a studio that generates words on your behalf, stores what you've written, and shows you what six well-read critics might think — all of which is plainly visible while it happens. The advisors are a lens, not a council. The model is doing the writing, and you are doing the judging. We log enough to keep it running, and not a byte more.
Author
