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.



