I'm an author, hands-on builder, and AI Craft leader from New Zealand. I use Codex and other AI tools to turn ambitious ideas into working projects, then study what makes AI-assisted development useful, reliable, and easy to understand.

I publish plain-English guides and small runnable examples for people who want to direct serious builds without giving up control of the decisions that matter. The work focuses on reliable agents, local AI assistants, human-reviewed workflows, source boundaries, scoped authority, and verification evidence. Most examples use synthetic data and make no model call.

• Build one useful private workflow now. Preview the Reliable AI Work Starter, then create a private copy for one recurring workflow with named sources, approval boundaries, durable state, and a reviewable handoff. Setup takes about 10 minutes and needs no app, model API, or public data.
• Learn how to direct and review AI work. Start with Build with Codex: A Plain-English Handbook to turn an idea into a bounded project and review the result without needing to read all the code.
• Find the right reliability pattern for your problem. The Toolkit Navigator recommends a public example by problem, experience, runtime, proof, and limitation. If you are reviewing AI-assisted code, begin with EvidenceGate.

These routes are designed for self-service. Each technical project names its runtime, first check, proof boundary, and limitations before you need to clone or install anything. Most use synthetic data, make no model call, and require no account beyond GitHub.

• Merged contribution: Skillware PR #263 added documentation for OpenAI-compatible model hosts and merged after the maintainer's requested revisions.
• Review contribution: On GitHub Docs PR #45193, I identified three bounded corrections to workflow-run status guidance; the author implemented all three before the pull request returned to maintainer review.

• Reviewable agent work. Context Boundary Examples catches answers that outrun supplied evidence, while Agent Action Authority Examples tests whether an approval still matches the exact action being proposed.
• Local-first assistant reliability. The Local Model Reliability Example keeps structured model output behind deterministic validation, and the Context Contract Compiler protects required context, explains selection decisions, and rejects stale receipts after declared material changes.
• Repeatable project confidence. Earned Confidence treats unknown as a state, not zero, and preserves what was known when a decision was made. The Green-Spine QA Pattern turns a representative journey into one memorable checkpoint, while the Public Repo Safety Kit helps keep private context out of public repositories.
• Human-directed game development. The Game Project Instructions for Coding Agents protects assets and saves, defines approval gates and checks, and preserves the human playtest across Godot, Unity, Unreal, or a custom engine.

These projects favour small synthetic examples, explicit limitations, and evidence a reviewer can inspect. A passing check is useful evidence, not a claim that the work is automatically correct, safe, or ready to publish.

Use the Lab's two-minute Toolkit Navigator to get one recommendation, or scan the complete toolkit map when you want every option and trust boundary at once. Agents and tools can read the same catalog from the machine-readable toolkit index.

I keep the broader lessons in Field Notes From Building With AI: what the surrounding harness should own, why useful work and permitted work are different, how to test boundaries before calling a model, and where human judgement still has to decide the outcome.

Away from the technical work, I write stories where fantasy, science fiction, and the apocalypse collide. I'm the author of The Mana Influx Series and Soul Spark Reclaimer. You can meet The Mana Influx Series on Amazon.

• Make capability accessible. People should be able to direct ambitious projects in ordinary language and understand the important decisions.
• Keep humans responsible. Models can propose and produce; people retain authority over consequential actions and protected areas.
• Leave something inspectable. Useful AI-assisted work has clear limits, relevant checks, honest uncertainty, and a handoff another person can review.

I like making real things, learning from where they break, and sharing the smallest useful version with other people.