Hi, I’m Nick Foscarini.
I direct AI agents and own what ships
I build software with Claude Code and Codex. I write the plan, keep each change small enough to review, and check every diff with tests and a careful read before it merges.
Markham, Ontario, Canada (opens in a new tab)Canadian citizen

Workflow
How I work with agents
Every change goes through the same loop: a written plan, small tasks an agent can finish, and a review before anything merges.
Plan First
I write a plan with the files, names, and tests for each step, so the agent builds what I decided, not what it guessed.
Small Tasks
Each task is one concern with its own tests, small enough that I can read the whole diff in one sitting.
Review Everything
I review each diff for correctness and fit, run the linters and tests, and check UI changes in a real browser.
Setup
How my agents are set up
Each part of my Claude Code setup fixes a way agents go wrong.
Context7
Loads the current docs for each library, so the agent works from the version the project uses, not from old training data.
Playwright
Opens each page in a real browser and screenshots it at phone and desktop widths, so UI changes are checked on screen.
Serena
Finds code by symbol and reference, so the agent reads the code it’s changing instead of guessing names.
Rules and skills
The repo holds its own rules and skills, so every session follows the same conventions without being told again.
Hooks
A hook installs the packages when a session starts, so the agent never works against missing or stale dependencies.
Where agents go wrong
- Guessing. Agents fill gaps with plausible guesses, so I give them the exact names, types, and values up front.
- Old APIs. Training data lags behind new library versions, so agents check the current docs before they write code.
- Passing tests. A test an agent wrote can pass for the wrong reason, so I watch it fail before I trust it.