Writing Decisions for a Future AI Assistant
I treat my Obsidian vault as long-term memory for my AI tools. When I write decisions down, an assistant can reconstruct what I did weeks later.
Honest takes on the tools I use daily. What earned its place in my stack, what I dropped, and why most "best of" lists are useless.
26 posts
I treat my Obsidian vault as long-term memory for my AI tools. When I write decisions down, an assistant can reconstruct what I did weeks later.
Exposing a stdio MCP server via HTTP on a small droplet, why a long secret URL ends up as the only authentication, and what tends to break first.
While setting up a new server, I found an API key in a nameless TextMate buffer and a base64-only password in an env file. This is how that happens, and what I changed.
15 days of Nginx logs: llms.txt was fetched 31 times, twice by an AI agent. And 4,334 of 4,565 "GPTBot" requests came from somewhere other than OpenAI.
I juggle a Mac Studio, a MacBook Air, and a T9 SSD by giving each type of file a single source of truth instead of trusting sync to do everything.
How I use Buffer to handle predictable updates for my one-person theme studio, what I still post by hand, and where I draw the line so social does not become a second job.
A practical look at what a small agency needs to make a Synology NAS searchable with a local AI stack on a Mac Studio, and what is still undecided.
A simple workflow for a five-team baseball club using Heja, iScore, WhatsApp, a 4-inning rotation PDF, and Instagram batching instead of heavy sports software.
A ChatGPT projects sync bug left my desktop app and the web client disagreeing about reality, so I treated one Mac as the source of truth and reached it remotely.
I cancelled my separate AI subscriptions and centered the stack on Littlebird, with a Mac Studio as a sovereign AI hub and a single context loop.
A client needed recurring reports from three different dashboards. One local n8n instance can take over the work, cut SaaS costs, and keep their data closer to home.
I stopped asking my automation to invent posts from topics and started feeding it five ugly bullets from real life instead. The system did not change much, but the writing did.
A practical line between “helpful AI code review” and “leaking your client’s business model,” based on operational privacy rather than legal fine print.
Photography got way better for me when I dropped the gear chase and treated images like components in a system: shot on an iPhone, finished in Lightroom, used on the web.
I killed CalendarBridge and RankPill, then audited my entire Setapp stack. This is how I trimmed the bloat without losing any real capability.
How I keep a Mac Studio and a MacBook Air in sync using Tower, Git and a few strict rules, so I can jump between desk and couch without merge hell.
Frontend dev writing about CSS, Make.com, biohacking, and building in public