Founders are surrounded by enthusiastic product feedback, but enthusiasm does not reveal budget authority or willingness to pay. A Thousand Interviews Overnight runs a deliberately sceptical synthetic customer panel locally through Codex, interviews each persona in isolation, and turns structured answers into an evidence map by morning.
The dashboard leads with “Who’d actually pay?”, an explainable segment ranking built from budget authority, price compatibility, verdict, pain, and Sean Ellis evidence. It also shows Van Westendorp acceptable price ranges, objections, a structured-signal map, complete transcripts, spoken brutal and enthusiastic quotes, and a one-click Markdown report.
The product is intentionally anti-sycophantic. Most sampled personas are lukewarm; sceptics and explicit non-buyers are required. A fresh deliberately bad-product evaluation produced 10 rejections from 10 interviews. The two demo studies are similarly sober: across 100 respondents they produced no automatic “buy” verdicts and surfaced the smaller trial-ready segments instead.
Generation happens only on the user’s Windows machine through spawned codex exec --json calls using the existing ChatGPT/Codex sign-in. Every persona is checkpointed immediately, calls are capped, rate limits pause and retry, and interrupted studies resume. There is no API key, API billing, embeddings endpoint, hosted generation, or server-side credential. The public Cloudflare Pages site is a static read-only viewer with two sanitized 50-person studies.
The product owner set the thesis, constraints, dashboard requirements, and transcript-realism gates. Codex converted those decisions into the structured contract, anti-sycophancy prompts, checkpoint/resume runner, deterministic k-means/PCA and Van Westendorp analysis, dashboard, tests, and documentation. GPT-5.6 powers the working local product: one call creates the weighted roster, one isolated call interviews each persona, and one evidence-bounded call names fixed locally computed segments. All commercial math remains deterministic and inspectable.
Synthetic interviews are fast hypothesis stress-testing, not real customer validation or observed demand. The output is designed to sharpen the next real interview, experiment, and pricing decision.
Built With
- cloudflare-pages
- codex
- gpt-5.6
- next.js
- react
- typescript
Log in or sign up for Devpost to join the conversation.