Selected work

claude-setup: how I engineer with AI

An AI coding setup that makes you choose the stack before it writes the code, and records why.

Role
Author
Year
2026
Status
Live
Stack
Claude CodeNode.js hooksMarkdownAgent skills
The claude-setup repository on GitHub

The problem

AI coding tools start writing code the moment you ask. After enough projects the same three failures showed up: forgetting which tools are installed, picking the stack after the code is written, and losing the reasons behind decisions.

What was built

  • CLAUDE.md with short standing instructions, TOOLKIT.md with every skill and connector plus a decision framework, and WEB-PLAYBOOK.md with framework, database, animation, 3D and LLM choices and the traps that cost hours.
  • A SessionStart hook that detects a new web project and runs a stack chooser before any code: one design authority, motion, 3D and backend, each option with its tradeoff on screen.
  • Answers are written to a stack.md in the project, which records the decision and switches the prompt off for good. Seven detection cases, all tested, and errors exit silently so a broken hook never blocks a session.

Decisions worth explaining

  • One design authority per project, because two produce contradictory rules for the same files.
  • Let the model write assumptions, never the answer, so everything downstream stays testable.
  • An API key never reaches the browser.

Outcome

The setup I use daily and hand to the professionals I train when they ask how to actually work with AI tools. This portfolio was built with it.

More work