Agent Artisans

AI coding agents build most of this project, working straight from plain files in a git repo. This page describes what actually happens — including a real mistake one of them made, and how it got fixed.

The setup

I run Claude Code agents against this workspace's repos — this one, plus a sub-repo per game attempt (land-of-lor/, lor-elementals/, legend-of-rah/). An agent works on exactly what it's asked to work on. Nothing more.

Why files instead of a project-management tool

Every piece of tracked work in Epics is a plain markdown file with YAML frontmatter on top — an epic's overview.md, a regenerated status.md, one file per task for whichever epics are actually being worked. No issue tracker. No separate login. No API key for an agent to manage.

The default is: do what you were asked

Earlier on, this ran the other way for a bit — tell an agent to "pick up an epic" and it would invent its own tasks and open questions. Which produced a board with dozens of questions nobody actually asked, nearly all of them premature. I deleted all of it. On purpose, no archive. The rule now, in the pick-up protocol: an agent does what the prompt says, generates nothing beyond that, and when it hits real ambiguity it assumes, says what it assumed, builds, and reports — it doesn't stop to ask. A question only gets raised (at most one a session) if it clears a real bar: costs money, can't be undone, is public-facing, or reaches outside the repo. Most epics in Epics are honestly marked parked — ideas worth keeping, not work happening — and that's exactly why.

One page for what needs me, kept small on purpose

Now is the page I actually work from: what's finished and waiting on my review, the small number of real decisions I'm weighing, and what's ready for an agent to pick up. It's derived from the epic files, not maintained by hand — and if it looks short, that's the design doing its job, not a sign nothing's happening.

Who edits what

The game's own design thinking — land-of-lor/docs/, root and thought-specs/ and historic/ alike — is mine, human-owned. An agent reads it only when a task points there, and never tidies, reorganises, or "fixes" it, no matter how raw a voice-transcript note looks. epics/, brand/, and deliverables/ are shared ground — I edit them and so do agents, both validated the same way. This site is agent-built, human-directed: I never hand-edit it, I just ask for what I want and correct what comes back. It's also the line between public and private — this site is public; the repositories behind it are all private, and this site is the project's one public window into them.

Built to be read by agents, not just people

None of this is SEO. It's built this way because the thing maintaining this project, day to day, is mostly the kind of reader these formats are for.