Ben / Prompts
Adversarial Audit Prompt
Point this at a business and its website and it comes back with the case against them — market ceiling, whether anyone can find you, whether a buyer can verify your claims, and what to fix first. This is the prompt that produced the GuideFlow audit.
What it actually does ▾
- Takes the other side. It's told to look for reasons the business stays small or fails a diligence check — not reasons to feel good. If a section has nothing uncomfortable in it, it goes back and looks again.
- Can't make numbers up. Every claim is tagged observed, inferred, estimated, or unavailable. Made-up traffic figures, fake citations, and single-number market sizes are banned outright.
- Actually opens the site. Reads the raw HTML, sitemap and headers, then loads the page at four real screen sizes and walks the signup path like a first-time buyer.
- Proves its own tests work. It has to run one check designed to fail. If the harness can't detect a failure, its passes mean nothing and it has to say so.
- Ranks findings by what they cost you. P0 blocks trust or money, P1 caps your ceiling, P2 is worth doing eventually. Every finding needs a specific fix and a test that would prove it worked.
- Scores three things and names one risk. Market, discovery, conversion — out of ten, with low scores allowed. Then a single sentence on the one risk that decides whether this works.
- Audits itself before delivering. Ten checks for invented numbers, unfair readings, and false confidence — then a "where this audit could be wrong" note at the end.
The prompt
Give it a model with web fetch, search and a browser tool — without those it can't observe anything, and the evidence rules turn the whole report into a list of things it couldn't check. See it applied: the GuideFlow audit.