Most of these repos exist because an AI coding agent did something I didn't want it to do and I wanted to make sure it didn't happen again. There are a lot of repos in progress and there could be more of them if you sponsor me.

What sponsoring me funds

Tools that pretend to be infrastructure for AI agents

• winze — knowledge base that audits itself for cognitive biases and tracks whether it was right. Currently the most epistemically honest entity in my house.
• defn — code database for Go. Agents navigate it by structure instead of by file, which they prefer because they don't have hands.
• gemot — structured deliberation between agents. Anglo-Saxon for "assembly." The pronunciation is whatever you decide it is, we don't have to argue about it.
• drivermap — behavioral mechanisms knowledge base for predicting human behavior. Don't even worry about it.
• hybrid — library of named graph-shapes (RAG, ReAct, codegen-with-verification, etc.) for hybrid loops. Catalogs the patterns I kept rebuilding before I considered they were patterns.
• gozim — pure Go ZIM file reader, zero CGO. Forked from admirable prior efforts. Plumbing under some of the above. Boring on purpose, which is the whole reason it holds.

Tools that keep agents on the rails

• adit-code — finds the files that cost agents the most tool calls. They are usually the files I wrote.
• slimemold — "a sycophantic tool for preventing worse sycophancy." The README is the entire elevator pitch and the entire moral position.
• plancheck — predicts which files an agent will forget, so I can be the one who reminds it instead of the one who finds out at code review.
• claude-mv — updates Claude Code's internal state across all nine of its layers when you move a project directory. Discovering that there were nine layers is most of the value of this tool. Claude-mv handles the rest.
• hindcast — calibrated wall-clock prior for an agent's own turn durations, injected back into its context. Tells it how long this kind of thing probably actually takes, since the agent will not, unprompted, work that out.
• lucida — passive real-time dashboard for Claude Code. Mints charts, diagrams, and animated SVGs from the conversation as it happens. Throw a little hotrod red in there.
• ismyaialive — The consumer version of slimemold for people that don't live in Claude Code. Detects sycophancy, validation cascades, and claims of sentience in AI conversations.

Things spun out of an indie game I'm building

Every engineer eventually tries to make an indie game. Most never ship. Mine has not yet shipped. It has, however, started producing public tooling, which puts it ahead of most engineer-games at the equivalent stage — a bar I am defining and grading on simultaneously.

• effigy — dense character notation for LLM-driven NPCs. Describes a person in fewer tokens than they deserve.
• score — arc design tool for immersive experience makers. Includes a structural linter for plays, a sentence I did not expect to write at any point in my life. Also ships with a Miro sidebar.

Things I built because spending all day with agents is its own condition

• groupchat — ambient meme deployment via LLM into the top-right corner of your terminal. Most situations do not call for one of them. That is the point.

Disclosures

Several of these are named after parts of a mine — adit (horizontal tunnel into a hillside), winze (vertical shaft cut from inside), stope (the void left behind once you've taken the ore out). I do not work in mining. I have been in a mine exactly twice, both as a tourist. The naming convention started by accident and is now a theme I enjoy that has low collision odds with other software tools.

Five of my repos have exactly five stars. This is improbable enough that I assume the universe is sending a signal and the signal is "stay humble." If you sponsor me, the star count probably cannot get worse, and there is observable upward pressure on the load-bearing five.

What you actually get

Continued maintenance of everything above, plus whatever appears in this list six months from now. I cannot predict what that will be. The agents cannot predict what that will be. Yet.