Roast your Twitter profile with AI-generated Severance wellness memes.
The work is mysterious and important.
Praise Kier.
lol but seriously though, with the release of GPT-4.5, I was really curious to see how good the latest LLMs were at generating subtle, personalized humor.
After a lot of iterating and testing with every frontier LLM, the results were... mixed... except for GPT-4.5, which was able to generate some genuinely funny and surprisingly nuanced jokes. In my experience, `gpt-4.5-preview` generates really solid results about 50% of the time, so in a batch of 10 memes, there's generally a few standouts.
Feel free to send feedback by replying publicly to this Twitter/X thread. You can also DM me, but I am more likely to reply if you are following me and the conversation is public.
I prefer public convos so the answers can help other people who might have similar questions, but if it's related to billing, please DM me.
- Post-MVP
- add sitemap
- improve landing page hero section
- improve header section + mobile support
- add basic pricing page w/ explanation for why I had to add the paywall
- add url resolving to
resolveTwitterUserto give more context - show generated date and model on fact detail page
- consider switching from prisma to drizzle
- consider switching from supabase postgres
- allow paying for more tweets + generations
- track third-party fact url access and image access for stats
The workflow to generate memes for a user:
- Fetch their twitter profile and 200 of their most recent tweets.
- Use `gpt-4.5-preview` to generate a list of text-based wellness facts based on a long prompt, some examples to guide the model, and the user's unfurled tweets as an array of plain text.
- Use `gpt-4o-mini` to try and extract the user's full first and last name, so we can call them by their Severed Lumon name (eg, Helly R).
- Use Vercel's dynamic OG image generation (which is powered by Satori under the hood) to generate the severance meme images on-the-fly.
MIT © Travis Fischer