dryrun

The problem

Interviews terrify capable people, and the fear is the thing that sinks them. What loses an interview is rarely not knowing the answer. It is freezing, rambling, going blank, panicking in the silence. Those are stress responses, and the only way to train them down is exposure.

Existing prep trains the wrong thing. A question list gives you all the time in the world to craft a perfect reply, which rehearses essay-writing, not interviewing. The AI tools that score you grade your delivery and confidence while missing whether your answer was any good, so you walk in feeling prepared and still bomb. And a whole shady wing of the market helps you cheat live during the real interview, which is reputationally toxic and increasingly banned by employers.

Who it is for

Anyone facing an interview they are scared of. Career changers, new graduates, people who have been out of the market, non-native speakers, the simply anxious. Any role, not just technical, because a pasted job link drives the tailoring. The nervous, inexperienced candidate is the core user. Everyone courts them; nobody serves them well.

The vision

A safe place to fail at interviews until they stop scaring you.

A realistic interviewer that puts you under genuine but controlled pressure (a ticking clock, a nudge when you stall, follow-ups that dig), and that never breaks character to correct you mid-stream. When you struggle it gives you a chance to recover. When you keep struggling it wraps up gently, the way a real interviewer would, and never tells you that you were bad. Then, when you are ready, it sits down with you and goes through your answers one by one, rewriting the weak ones together, in your own words.

You climb a difficulty ladder as your tolerance grows: text first (low fear, easy to start), then voice, then video. The medium is the difficulty dial. You come back to climb it, and you watch yourself stop being scared.

You leave not with a score, but with better answers you wrote yourself and a nervous system that has already been there.

Why us, why now

The raw capability is one prompt away. That means this cannot win on the model or the questions. It wins on product design and coaching quality, specifically two things almost nobody has shipped: a realistic interview that does not correct you, and a collaborative rewrite of your answers afterward. The only place those two exist together today is a hobbyist prompt, not a product.

It also wins by being on the right side of two lines. We help before and after the real interview, never during it. That is the ethical line that separates legitimate preparation from the cheating tools employers are now banning, and it is also the moat, because the trustworthy positioning is the durable one.

What winning looks like

Someone walks into a real interview calm, because they have already done twenty harder ones this week. The answers they give are answers they rewrote themselves a few days earlier. The fear that used to make them avoid applying is gone. They got the job, and they did it on their own merits, sharpened.

The press release is written as if the product has already launched. It is aspirational, not a record. The date and numbers are placeholders to force clarity, not commitments. The FAQ is where the hard questions get answered honestly.

Press release

dryrun helps nervous candidates practice interviews until they stop being scary

Paste a job link, get interviewed for real, then rewrite your weak answers together. No scores, no coaching during the real thing.

FALL 2026 - Most interview-prep tools hand you a list of questions or grade how confident you sounded. Neither fixes the thing that actually loses interviews: freezing, rambling, and panicking under pressure. dryrun takes a different approach. It runs a realistic mock interview tailored to the exact job you are applying for, never interrupts to correct you, and only afterward sits down to improve your answers with you, one at a time.

You start by pasting the link to a job posting. dryrun reads the role and generates the questions you are most likely to face. Then the interview begins. There is a clock. If you stall, it nudges you, the way a real interviewer's silence does. If an answer is weak, it digs with a follow-up rather than moving on, giving you a chance to recover. Strong answers earn a longer, deeper interview. If you keep struggling, it wraps up gently and professionally, never telling you that you did badly.

When you are ready, and not a moment before, the coaching begins. dryrun walks through your answers with you and helps you rewrite the weak ones in your own words. You leave with answers you actually believe, and with the experience of having already been under pressure.

"I used to avoid applying because the interview scared me more than the job excited me," said an early user. "After a week of dryrun I walked into the real one calm. I had already done harder versions of it."

dryrun deliberately helps only before and after the real interview, never during it. "There is a whole category of tools that whisper answers to you live, and employers are banning them," said founder Anthony Rasch. "We are the opposite. We make you better beforehand so you do not need help on the day."

dryrun is available now, starting in text. A single one-time purchase includes a generous bundle of full mock interviews. Voice practice is rolling out next.

FAQ

Why would I use this instead of just asking ChatGPT to interview me? You can, and it is the right way to test the idea, which is exactly how this product started. But a blank chat gives you all the time in the world, which trains essay-writing, not interviewing. dryrun adds the things a chat will not: real time pressure, a nudge when you stall, an interviewer that reacts to hesitation, a structured rewrite of your answers, and a record of your progress so you can watch yourself improve. The value is the method and the pressure, not the model.

How is this different from Final Round AI, Yoodli, Big Interview, Huru, and the rest? Two things they do not do. First, they correct or score you after every answer, which breaks the realism. dryrun stays in character through the whole interview and saves all feedback for a dedicated debrief. Second, they hand you a "model answer" to copy. dryrun rewrites your answer with you, so what you walk away with is yours. Job-link tailoring and follow-up questions are now table stakes, so those are not the pitch.

Isn't this just another interview-cheating tool? No, and the distinction is the whole brand. The cheating tools (Cluely, the live "interview copilots") feed you answers during the real interview. Employers including Amazon and Anthropic now explicitly ban them. dryrun does the opposite: all help happens before and after, never during. We make you genuinely better, we do not help you fake it.

What does it cost you to run, and what do you charge? A full text interview plus coaching costs roughly 15 to 50 cents in AI usage, lower with prompt caching and model-tiering. Interview prep is bursty and self-terminating (people prep for a few weeks, get hired, and leave), so even a heavy user's lifetime cost is single-digit dollars. That is what makes a one-time price viable. The plan is a one-time purchase with a generous included bundle of interviews (feels unlimited to a normal user, caps our downside on outliers), with optional credit top-ups. Voice, which is several times more expensive to run, comes as a higher tier.

What happens if I panic and can't answer? One bad answer does not end anything. dryrun digs once, gently, to give you a chance to recover. Only after a repeated pattern of struggling does it wrap up, and when it does, it is always polite and professional ("thanks for your time, we will be in touch"), never "you failed." Then it lets you rest, and the coaching starts whenever you feel ready.

How do you tell a panicking candidate from one who is just thinking? This is the hardest part, especially in text where the signals are thin (short answers, "I don't know", long pauses before typing, rambling, asking to skip). The product deliberately errs toward staying in the interview. A false "you're struggling, let's stop" aimed at someone who was actually fine is the worst experience we can create, so when unsure, we continue.

Why text first instead of voice, since interviews are spoken? Activation. A scared user will start by typing long before they will start talking to a robot. Text lowers the fear enough to get them in, where they see the value, get comfortable, and then graduate to voice. Text is also far cheaper to run, which protects the economics early. Voice and video come later, as higher rungs on a difficulty ladder.

What stops a competitor from copying this? Nothing stops the obvious version, which is already a crowd. The defensible part is execution quality of the rewrite loop (genuinely hard, which is why nobody ships it well) and the trust of the legitimate-preparation positioning. The moat is craft and brand, not technology.

Why one-time instead of a subscription? The audience is often unemployed and churn-prone, and the loudest emotional complaint about the incumbents is subscription billing rage (surprise renewals, cancellation hell). A one-time price avoids that and matches a use case that naturally ends when the user gets hired. If we ever offer a subscription, effortless cancellation is the differentiator, not a trap.

The engineering "how". This is a living design at idea stage; nothing here is locked until we decide to build. Scope is the text-first MVP unless noted.

Principles

• Two phases, one wall between them. The interview never corrects you. All feedback lives in the debrief. This separation is a product law, not a preference.
• The rewrite loop is the moat. Spend quality (and the better model) here. A shallow "here is a model answer" collapses us into the commodity pack.
• Bias toward staying in. Every adaptive decision (wrap up, flag panic) defaults to the gentler, more forgiving option when uncertain. A false "you're struggling" is the worst experience we can produce.
• Cheap to run in text. Model-tiering and prompt caching are first-class, not optimizations bolted on later.

Flow

job link
  -> role scan + question generation
  -> interview (no correction, time pressure, adaptive length)
  -> gentle wrap-up
  -> [user rests, returns when ready]
  -> debrief: per-answer review + collaborative rewrite

1. Intake: job link to questions

Take a pasted job posting (URL or text). Extract role, seniority, domain, and likely competencies. Generate a tailored question set: behavioral, motivational, and verbal-technical ("walk me through how you'd design X", "explain a tradeoff", "describe a hard bug"). No live coding in the MVP (different product, crowded sub-niche).

Resilience note: many job boards block scraping. Accept pasted text as the reliable path; treat URL fetch as best-effort with paste fallback.

2. The interview

• One question at a time. Never correct, score, or coach mid-interview.
• Time pressure: a visible clock per answer. A nudge fires if the user stalls ("I'd like to hear your initial thoughts, even if they're rough"). Text-with-pressure, not a cozy editor with unlimited time.
• Adaptive depth: strong answers earn a follow-up that digs deeper; the interview runs longer. This is realism, and interview length becomes an implicit, unspoken score that feeds the progress system.
• Recovery on a weak answer: dig once, gently, before concluding anything. One bad answer is not a bad candidate.
• Wrap-up: always gentle and professional, never "you were bad". A real interviewer does not say "sorry, ciao". The hard early-end only triggers on a repeated pattern of struggle (roughly two to three), and ends with the realistic, dignified "thanks for your time, we'll be in touch".

3. Struggle / panic detection (the hardest piece)

Text signals available: very short answers, explicit "I don't know", long typing latency (measurable), incoherent rambling, requests to skip. None are reliable alone; panic, thinking, and natural terseness look alike.

Rules:

• Accumulate signals across answers; never end on a single event.
• The "dig once" follow-up doubles as the probe: if the gentle follow-up also collapses, that is a real pattern; if the user recovers, staying in was correct.
• When uncertain, continue. Conservative by design. We would rather over-stay than wrongly reject the person we built this for.
• Difficulty tier modulates aggressiveness: beginner mode rarely hard-ends; realistic mode will end early like a real interviewer.

4. The debrief (collaborative rewrite)

• User-paced. Starts only when the user chooses. The end of a rough interview is the churn cliff, so the off-ramp is a warm invitation ("take a breath; when you're ready, this is where it stops being scary"), and the session is saved so they can return.
• Walk answers one at a time. Ask the user how they think it went before judging (mirrors good human coaching).
• For weak answers, rewrite together in the user's own words, not a handed-down model answer.
• Per-phase gentleness: maximally gentle at the interview exit, candid-but-kind in the debrief. A gentle debrief that inflates the user ("you were great") makes us the tool they blame after they bomb the real one. Gentle goodbye, honest rewrite.

Difficulty ladder (medium roadmap)

The medium is the difficulty dial and the retention loop.

1. Text with pressure (MVP). Lowest fear, cheapest to run.
2. Voice. Felt pressure, silence detection in audio. Higher cost, higher tier.
3. Video. Eye-contact and body-language pressure (HireVue-style). Build only if demand shows up.

Users graduate up the ladder as tolerance grows. Climbing is itself the progress narrative.

Model strategy and cost

• Model-tiering: a cheaper model conducts the interview turns; the best model does the rewrite (the part that has to be excellent). Roughly halves cost while protecting the moat.
• Prompt caching on the re-read transcript context to cut input cost on every turn.
• Target: ~$0.15 to $0.50 per full text session; aim for the low end via the above.
• Budgeting: track tokens per session; enforce per-user usage caps (bundle / credits) so cost is bounded against a one-time price.

Data model (sketch)

• users (auth, entitlements, credit balance / bundle remaining)
• sessions (job context, difficulty tier, status, started/ended, length-as-score signal)
• transcripts (per-turn question/answer, timing signals for the detector)
• rewrites (original answer, collaborative final, per-answer)
• progress (cycles completed, ladder rung, return cadence)

Tech stack

Studio defaults: Next.js (App Router) on Vercel, Supabase (auth + Postgres). AI via the studio's preferred provider with caching. Defer final choices until build; keep the MVP a thin layer over the interview-and-debrief engine.

Open questions / risks

• Struggle detection accuracy in text is the top product risk. Start conservative; consider whether to keep it out of the very first cut entirely and add it once the core loop is validated.
• Rewrite quality is the moat and the hardest UX. Underbuilding it is the main way this fails.
• Job-board scraping reliability; paste-first mitigates.
• Pricing mechanics (bundle size, credit top-up UX) need a real number once we have session-cost data from the beta.

Cheapest first build

Before any of the above, ship the interview-and-debrief loop as a Claude skill / prompt the founder uses for their own job hunt. Zero infrastructure, validates the experience and the rewrite quality, and is the zeroth-step beta. The only comparable thing in existence today is exactly this: a hobbyist Claude skill.