← posts 
mirofish2026-05-14
# 26 synthetic personas pre-mortemed my portfolio
before my first real visitor, i ran 26 fake ones through the full funnel.
the tool is MiroFish. a multi-agent simulation engine. you seed it with your content, it constructs a digital world, runs LLM agents that interact and evolve across simulated time, then generates a report.
i used a pre-mortem framing. the prompt: 'It is November 2026. Guillermo's portfolio failed to generate the inbound he wanted. Walk through what you saw, in order, and explain why you didn't act.'
26 personas. hiring managers, technical recruiters, Murex specialists, peers. 72 rounds of interaction covering 74 emulated weeks. under $1 in API calls. one afternoon.
mirofish, #1 github repository of the day
## what they found
93% convergence on one finding: the LinkedIn headline was too vague. not the portfolio. LinkedIn.
the page that came closest to convincing them: /work. stopped by two things: missing outcome metrics (57%) and vague project descriptions (36%).
the mirofish report, with the agent stream running
## one caveat
MiroFish hallucinated a project name in the specific evidence. the failure-mode shape was real. the receipts were confabulated. synthetic users are a pre-screening tool, not real traffic substitutes.
## what i did with it
took the top 3 findings to Claude Code. fixed the gaps.
then Claude Design to keep the design system aligned after the content changes.
one thing worth knowing if you use it: the design system export takes up 27.5% of your context window. not a dealbreaker, but it's not free. if you're running complex sessions on top of it, plan for it.
## where it actually helps
anywhere you need a brutally honest pre-mortem before you have real traffic.
landing pages before you spend on ads. product pitches before you send the deck. intake scripts before a client reads them for the first time. messaging before you commit to a positioning. anything where you're guessing how a cold visitor reads you.
it's not a substitute for real user interviews or live A/B tests. it's what you run upstream of those, when you don't yet have users to test with.
## will i use it again
yes, with the right framing.
the question to ask MiroFish is not 'which version converts better' but 'what are the ways this could fail.' it answers the second question well. trust the failure-mode shape. don't trust the specific receipts.
for any project where i'm launching something cold into a new audience, this is now part of the pre-launch checklist.