Build an AI brain for your dev work

A new session opens and you are back at zero: re-explain the project, re-explain the conventions, then watch the AI repeat the mistake you corrected last week.

I run my studio on a small set of plain text files that the AI reads at the start of every session. Below is the whole setup, with templates you can paste today.

The three layers

1. Instructions — a CLAUDE.md file that says who you are, how you work, and what the tool must never do. It is read automatically at every session start.
2. Memory — one file per fact the AI should keep across sessions: project state, decisions, things you corrected it on.
3. Skills — reusable playbooks for tasks you repeat, such as your deploy process or your review checklist.

All of it is plain markdown, versioned in git. No vendor lock-in, no database.

Layer 1: the instructions file

Create a CLAUDE.md at the root of your project (or ~/.claude/CLAUDE.md for global rules). Keep it short and blunt. Mine enforces rules such as verifying before claiming done and never committing unrelated files.

Copy-paste starter:

# Working rules

## Verification — non-negotiable
- Never report a task complete until typecheck and lint pass.
- A file write is not success. Code must compile.

## Git discipline
- One feature per branch. Scoped commits: add exact files, never `git add .`
- Never push without running the project checklist.

## Output discipline
- Lead with the result. No filler, no restating the question.
- When the answer is one line, give one line.

## Project facts
- Stack: [your stack]
- Commands: typecheck `npx tsc --noEmit`, lint `npx eslint . --quiet`, test `npm test`

The last section matters most because it gives the AI commands to check its own work. In my experience that single block removes half the babysitting.

Layer 2: memory files

One file per fact, plus an index. The index is loaded every session, and individual memories are pulled in when relevant:

memory/
  MEMORY.md            <- index: one line per memory
  api-rate-limits.md   <- one fact
  client-x-scope.md    <- one fact

Each memory file:

---
name: api-rate-limits
description: Our provider caps at 100 req/min — batch accordingly
type: project
---

The payment provider rate-limits at 100 requests/minute.
Why: hit this in production on 2026-05-02.
How to apply: batch webhook retries, never loop unthrottled calls.

The rule that makes this work: after any correction, write a memory. If you told the AI "no, we use tabs here" once, write that sentence into a file so you never have to say it again.

Layer 3: skills

A skill is a markdown playbook the AI loads when a task matches. Anything you would otherwise re-explain, from design conventions to deployment steps, fits here.

---
name: design-taste
description: Invoke before any UI work — typography, spacing, motion rules
---

## Typography
- Body text max 65 characters per line
- Bold for emphasis, never italic. Underline only for links.

## Motion
- Animate only transform and opacity. Keep under 300ms.
- Always honor prefers-reduced-motion.

Claude Code picks these up from .claude/skills/ automatically. Ten short skills beat one giant document: the AI loads only what the task needs.

Subagents: the multiplier

On a big task, have the main session dispatch subagents instead of doing everything itself: one explores the codebase, one implements, one reviews. Each gets a clean context and a single job, while the main session orchestrates and keeps the overview.

Practical rule: research and exploration go to subagents, decisions stay with you and the main session.

Start today, small

1. Write a 20-line CLAUDE.md with your verification commands.
2. Create memory/MEMORY.md, empty.
3. Next time you correct your AI, save the correction as a memory file.

Step 3 is the loop that matters: correct once, write it down. Add skills and subagents later, when the same correction or the same task comes back.