The Broader Practice
Work
The author's broader practice spans creative and engineering work, all directed by one premise: AI is a tool the craftsman points, not a partner who decides.
AI research and development
A deliberate exploration of how AI coding tools, particularly Claude Code, can compound across sessions. The work plays out across three working examples:
Long-form fiction work
Novel writing with persistent worldbuilding. The challenge is keeping the AI coherent across a multi-book series, disciplined about voice, canon, and continuity. The Shattered Covenant is the working example.
Personal website development
An evolving content-and-design surface that exercises AI assistance end-to-end: layout, content systems, build and deploy pipelines, and the trade-offs of choosing a hosting platform. This site is the working example.
Interactive narrative game development
Porting a substantial prototype into a shipping product. The discipline here is scope: avoiding unnecessary refactors, verifying work before declaring it done, and steady multi-session execution against a fixed plan.
Across all three, the question is the same: what configuration of tools and instructions actually makes Claude Code reliably better at the work, and which features should be cut because they don't earn their keep?
Working principles
If AI is a horse under harness, these are the four pieces of tack. Each one exists because something in the work demanded it; each one comes off when the work no longer does.
- Earn the rules. Don't load the AI with instructions before something has gone wrong that needs one, and retire a rule when whatever caused it is gone. Absence counts too: if the same correction keeps coming back because nothing prevents it, that absence has earned a rule.
- Strict at the seams. Where the AI's output has to be exact, check it hard. Where it is still thinking, give it room. Tight at the edges, loose in the middle.
- Work with the model, not against it. When tempted to add one more rule telling the AI what not to do, look for what is drawing the behavior out and remove that instead. Telling it "don't" rarely sticks. Treat causes, not symptoms.
- Cite when proposing. When the AI suggests a change, it has to point to a decision I have already written down, not a story it invented in the moment to sound convincing. Judgment stays anchored to the record.
Every feature has to prove its worth or get removed. A good session isn't measured by how much was added, it's measured by whether the work is better than when it started.