https://ajpaschka.github.io/gggggit/ I started getting some cold emails of people wanting to put their favorite get repos on the list ——so I spent a few minutes elevating the site today. so if you have favorite repos, please submit them and I will include them in the list.

Inspired by ffffoound and wearehunted websites from back when the internet was fun. https://ajpaschka.github.io/gggggit/ It's fun to materialize an idea and get it shipped in under an hour.

The workflow: Hermes (orchestrator) routes every request. Skills already declare which model tier they need (the model: frontmatter is real, not decorative). When a request touches sensitive data, Hermes defaults to the local model — Gemma, via the Ollama instance Hermes already runs, currently idle — instead of escalating straight to Claude. 1. Local model scopes the request — reads the sensitive material, decides what needs coding. It only proposes; it never makes the final call. 2. A deterministic script does the actual encoding — swaps real names/values for stable per-entity codes via a codebook (table lookup, not AI judgment). . 3. A preflight check verifies nothing raw survived before anything leaves the machine — the check that can demonstrably fail on a bad run, not just a claim that coding happened. 4. Only then does it ask permission to send the coded payload out — your "greenlight and asks permission" step. 5. Claude works entirely on coded tokens — never sees a real name, real entity, real number. 6. The same local codebook decodes Claude's answer back to real names before you see it. That's the full round trip. Claude never sees a real name; you never see a code. What am I not thinking of?

Every build has to have a manual for operation built alongside it. I’m shipping enough stuff that a plain language ‘how to use’ this document is now standard for me. Each draft of something needs to be understood by the client. A simple thing that is painfully obvious in retrospect.

@Eugene Seva and I jumped on a video meet after I commented on one of his posts. He shared the GitHub repo of his work with me. A cool build of a Japanese digital business card for realtors. QR code, AI chat, mortgage calculator. Live site, paying customers, 140 merged pull requests, still shipping daily. Good build. I spent the day trying to work out the Western version. Never got there. What I got instead was better. That product works because the company behind it is a Tokyo brokerage sitting on 4.5 million sales records and 6,000 new listings a day. The app is a skin over a filing cabinet they already owned. Take the data away and it's a link-in-bio page with a chatbot attached. In the US that data is MLS-licensed per organization, and you need a contract with each one even going through an aggregator. That's a funded-company problem, not a weekend one. So I went looking for something that runs on nothing licensed. Six shapes, six walls. Business cards lose to a free tier with 150,000 reviews. Valuation needs an AVM license. Chat and nurture is held by people charging $395 a month. The fifth one nearly got me. Since 2024, buyers sign a written agreement about agent compensation before touring a house, and everyone is confused by it. No tool exists. I had the whole thing designed. Then I checked whether a real estate agent is permitted to explain a contract to a client. They are not. That's unauthorized practice of law, in all fifty states, and it's the agent handing the document over who carries the exposure. Insurers started writing generative-AI exclusions in January, so it wouldn't be covered either. Here's the part you'll care about. The constraint ended up as the folder structure. "Never produce client-addressed output" is a rule, and a rule sitting in rules.md is a suggestion. So it's also a check script. It reads every skill file, and if a client salutation turns up in one, it exits 1. Same for listing data. Each skill declares its inputs in a table, and the check fails if any input is a source we can't legally touch.