Matt Holden
@holdenmatt
LLMs are fuzzy microprocessors. exploring plain english programming @ http://fuzzycomputer.com ex @dropbox, @google, 2x founder
how am i supposed to work if claude is down? am i supposed to write code and tests...by hand, like typing each character? that's like saying i should write assembly or punchcards because my C compiler is down
LLMs are fuzzy microprocessors for 2 years i've been thinking about this - what if we had fuzzy compilers? fuzzy MapReduce? fuzzy design systems? what if we could program in plain english? i’ll be exploring here: fuzzycomputer.com/about models are finally good enough to build…
Anyone know if it's helpful to document "commands" for AI tools like claude code? Anthropic's best practices suggest yes, so i put stuff like this in my CLAUDE.md. But shouldn't it be able to just read my package.json? (I wish I could add comments there instead)
I quite often add "no code, discuss first" when working with an LLM If I want feedback about an idea or draft, "be critical, not sycophantic" seems to help When working on a bug, I'll say "root cause before you fix" Any other weird incantations like this folks find useful?
I gave a fun/zany talk at my local AI Builders meetup Luckily @dalmaer wrote it up so I don't have to I'm quite excited about some of the underlying ideas here, so a lot more to come (soon)
Check out @holdenmatt’s vibe improv! I wrote about it here: ainativedev.io/news/improv-me…
Just had my first fully "spec-driven dev" experience, pretty interesting I built a little open source python lib + cli tool (will share more soon) I didn't write a single line of code. Instead I used Opus to iterate on READMEs, TESTING plans, and a few specs (describing edge…
i still don't know what .mjs or .cjs files are, and at this point i'm afraid to ask
i just closed ~200 browser tabs. i think some had been minimized for over 2y any fellow chaotic tab hoarders out there?
anyone know if there's a programmatic way to generate a @MeetGamma slide deck from a markdown outline (like the /create/paste feature, but via api)? something like a POST call or integration trigger or maybe n8n/gumloop thing? (cc @thatsjonsense )
yeah, +1 for retiring "prompt engineering" my "prompts" these days (for claude code) are getting super long and complex, more like composing entire react trees of components than writing a little inline js file "spec" is closer than "prompt" but also not quite right "model…
+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window…
giving a live demo tomorrow that needs claude code in order to work. in the claude app, I constantly get these, which doesn't inspire confidence in reliability. i'm using claude code with an API key. anything i should do to make it reliable? never used bedrock, vertex, or…
There are two competing narratives about AI: (1) there's too much hype (2) society is being too dismissive and complacent about AI progress. I think both have a kernel of truth. In fact, they feed off of each other. The key to the paradox is to recognize that going from AI…
a pattern i'm finding helpful is to have a "primary" and "secondary" LLM tool currently i'm using opus 4 (so good!) in the claude desktop app as my primary, and gpt-4.1 as my secondary in chatgpt (installed as a chrome PWA). that gives me two intelligences i can cmd/alt-tab…
One of my favorite @aiDotEngineer talks. What might the Bitter Lesson imply for UX design of AI products? As models have gotten better at reasoning and using tools, last gen UI starts to feel like it’s in my way. One reason Claude Code feels so good (as a power user) is it…
Not sure when my @aiDotEngineer talk will be posted, so for those who asked here's a link to my slides (couldn't add notes so if you anything needs explanation just reply here & I'll thread the context): figma.com/deck/0NURDonIV…