The Vault: My Complete AI-Engineering Setup

Every template, slash command, hook, and prompt I actually use — in one place, kept current.

I write a lot about how to work with coding agents. This post is the what: the exact files and configurations I use every day as a Data Science Lead shipping AI systems to production.

This is a living document. Whenever my setup evolves — new slash commands, better hooks, revised templates — I update this post and note it in the changelog at the bottom. Paid membership includes access to the Vault and its updates while your subscription is active.

Everything here is worth more than a year of the subscription on its own: it is the exact setup I use every day, kept current, so you skip the months of trial and error I paid for. Access to the latest configuration and the other paid benefits is included while your subscription is active.

What’s inside

1. My global CLAUDE.md template (the rules every project inherits)

2. My project-level CLAUDE.md template

3. My slash command library (PR creation, code review, spec interviews, verification)

4. My hooks (guardrails that run automatically, so I don’t rely on the model remembering)

5. My prompt patterns (few-shot scaffolds, verifiable-task templates)

6. My parallel-agent workflow (git worktrees + session layout)

7. My memory setup for continual learning

---

1. Global CLAUDE.md template

This file applies to every project on my machine. The principle: facts, invariants, and preferences only — recurring problems get mechanisms (hooks, scripts), not more prose rules the model can ignore.

# Global instructions

## Interaction style
- At the end of EVERY response, if you are asking me to do something or there is
  anything I need to test, present it as Markdown checkboxes — never buried in prose:
  - [ ] <task 1>
  - [ ] <task 2>
  If there are no tasks for me, say so explicitly.
- Below the tasks, end every response with a short recap: what we did, what is
  left for me, and whether I can close this thread.

## Effort preferences
- For non-trivial implementations, always take the high-effort path: plan first,
  spend more tokens, verify before reporting back. A slower implementation that
  works beats a fast one I have to spend an hour correcting.

## Git invariants
- Treat main and release branches as protected: no force-push or reset --hard without
  explicit per-push approval.
- After opening a PR, watch CI to completion — local hooks passing is necessary,
  not sufficient.

## Code style
- Variable names: full descriptive words, never abbreviations (loop indices exempt).
- Python: use 3.13 syntax, always write return types, never use the Any type.

## Environment truths
# Facts about YOUR machine the agent keeps getting wrong — these save the most
# time of anything in this file. Mine, as examples:
- ~/Documents, ~/Desktop, ~/Downloads are permission-blocked for shell tools on
  macOS: an empty ls means access denied, NOT an empty folder.
- Long package installs (brew install/upgrade) take 1–10 min: always run them
  in the background instead of blocking the session.
- Credentials live in env vars or ~/.config/<service>/credentials — never
  hardcode keys in source, markdown, or shell commands.

Why the checkbox + recap rules are the most valuable lines in this file: coding agents write a lot of text, and when an ask for you is buried in the middle of a long response, you will miss it — I did, repeatedly, and it quietly hurt quality. The checkboxes make every ask impossible to miss, and the instant recap means I can return to a thread after 10–30 minutes away (which happens constantly when running agents in parallel) and pick up the context in seconds. Claude Code has a built-in recap feature, but it is often delayed — prompting the agent to write its own recap gives it to you immediately, every time.

Why an effort preference belongs here: performance scales with tokens spent. The real choice is between a fast implementation in ten minutes plus ninety minutes of you correcting it, or forty minutes of high-effort agent work that is simply done. Encoding “always choose quality over speed” globally means you never have to remember to ask for it.

Paid subscribers also get the monthly deep dives with working code, every update to this setup as it evolves, and the full archive.