JAMIE

I spent a long time avoiding AI. Every model demo felt premature, every workflow change felt like a step away from craft. Then it kept coming up in job interviews — by the fourth conversation, I'd run out of polite ways to say I didn't have an AI workflow.

So I set out to learn how to work with AI well, not just use it. Which surfaced a problem: methodologies were everywhere, each with strong opinions, and most assumed I'd commit to one vendor, one framework, one company's playbook. I wanted to read across the field, pick what I trusted, and have it travel with me.

JAMIE is what came out — a personal synthesis. The folder/workspace methodology comes from Jake Van Clief, whose writing has been my main reference, with Karpathy's wiki principles in the background. The soul document and persona ideas come from OpenClaw. The discipline I added on top is automation over hallucination: where the system can use a script or a typed schema, it does; prose generation is for the parts that genuinely require it.

The ideas aren't original — the synthesis is. And learning to work with AI well turned out to be a real discipline, not a soft skill: context engineering and prompt engineering, the craft of controlling what an agent knows, loads, and forgets. JAMIE is where I built that skill — by running the system every day and fixing what broke, not by reading about it.

Markdown files and conventions, not an app. JAMIE is text I can grep through, not a dedicated runtime with its own UI. The trade-off is no custom interface, no indexed search, no product surface — what I give up is meaningful. The payoff is that the same system runs across Claude Code, Codex, and Gemini today — switching engines is a config change, not a rewrite — and next year's model inherits the whole thing intact. Building it as files made everything else possible.

Automation over hallucination. Where a script or a typed schema can do the work, use it; reserve prose generation for the parts that genuinely require it. The cost is more upfront engineering than "ask the agent to handle it." The payoff is that deterministic operations stay deterministic — the agent doesn't drift on counts, file paths, or naming. Hallucinations compound; structured boundaries don't.

Memory is distilled, not dumped. Working memory entries are curated artifacts — frontmatter-tagged, machine-indexed — not chat-log archives. Curation happens at retro time, manually. The labor is real. The payoff is that retrieval stays fast and signal density stays high. A memory system that only accumulates becomes a graveyard to wade through, which is worse than no memory at all.

Workspace-scoped context. The agent loads only the files the current workspace needs: a writing task pulls in writing conventions, a code task pulls in entirely different files. The cost is partitioning the hub into separate workspaces with their own context files, which is upfront design work. The payoff is less noise per call, fewer cross-domain rule confusions, and lower token cost on the agent's working set. Context engineering at the filesystem level.

Self-review before it ships. JAMIE reads its own output through named expert lenses — a senior engineer, a designer, a hiring manager — each grounded in a catalogue of distilled professional canon (Clean Code, Fowler, Nielsen, Strunk and White) rendered as detectable anti-patterns. A lens flags against its catalogue; it doesn't free-associate, and a lens with no catalogue behind it doesn't run, by rule. The cost is maintaining the canon. The payoff is review that's principled instead of impressionistic — the same artifact gets checked for maintainability, for clarity, and for hiring signal before I call it done.