The 80-Line CLAUDE.md Replacing a 4-Hour Claude Course

By Lazar Milicevic · Published July 12, 2026 · 7 min read

You bought the course. You installed Claude Code. Day one felt like magic. Day three it imported libraries not in your package.json, rewrote the auth handler when you asked for a typo fix, and generated a Prisma schema for tables that don't exist. That's not a demo trick failing — that's a missing file.

Every serious Claude Code setup lives or dies on a single markdown file at the repo root: CLAUDE.md. It's the instruction manual Claude reads before every response. No file, no context. No context, hallucination. I've been running the same 80-line version across three production products for 90 days — an invoicing platform, an internal ops bot, and a lead-gen tool. Different stacks, different clients, one file shape. Here's exactly what's in it and why the six lines that matter most are the entire course.

Why courses spend 4 hours and skip the one file that matters

Most Claude Code walkthroughs burn 3+ hours on install screens, VS Code themes, MCP server tours, and philosophical takes on why AI matters. Zero minutes on CLAUDE.md. The reason is simple: install steps are easy to film and easy to sell. A production instruction file is the output of failure — you only know what belongs in it after Claude has broken something you had to roll back.

That's why the file took me four rewrites over six months to land. Each forbidden line in my current version maps to a specific incident. Raw SQL migration in a Prisma repo? Rollback #1. Silent npm install of a validation library we don't use? Rollback #2. Auth handler rewritten on a "fix the typo in the error message" prompt? Rollback #3. The courses don't have the scars because they never shipped a production feature under a real client SLA.

Anthropic actually documents this pattern in the Claude Code memory docs — CLAUDE.md at repo root is loaded automatically into context. The documentation exists. Nobody teaches what to put in it.

The five sections of an 80-line CLAUDE.md

Five sections, no more. Every section earns its lines by preventing a specific hallucination class. Here's the actual skeleton I ship on every project:

# Project: Fakturko API

## 1. Context
Node backend serving an invoicing API for SMB clients.
Status: production, live customers, revenue-generating.
Audience: non-technical business owners via a React frontend.

## 2. Stack (pinned)
- Node 20.11 LTS
- TypeScript 5.4, strict mode
- Fastify 4.26
- Prisma 5.7 (migrations, not raw SQL)
- Postgres 15
- Vitest 1.4 for unit, Playwright 1.42 for e2e

## 3. Forbidden without explicit approval
- Never run `prisma migrate deploy` or `prisma db push`
- Never modify files in `/prisma/migrations/**`
- Never install a new npm dependency — ask first
- Never rename or remove env vars in `.env.example`
- Never touch `/src/billing/**` without a "human-reviewed" comment
- Never write raw SQL migrations — use Prisma schema only
- Never change `tsconfig.json` `strict` or `paths`
- If a change touches >3 files OR any forbidden path, stop and ask

## 4. Verify commands (exact)
- `npm run test:unit`
- `npm run test:e2e`
- `npm run lint:fix`
- `npm run typecheck`
- `npm run build`

## 5. Commit style
- Conventional Commits: `feat:`, `fix:`, `chore:`, `refactor:`
- Every message ends with `[TICKET-123]` or `[no-ticket]`
- One logical change per commit — no mixed refactors

That's it. Eighty lines with whitespace. No YAML frontmatter, no cute markdown tricks, no links to a Notion page Claude can't read.

Why each section exists

• Context sets the stakes so Claude treats a live billing repo differently than a weekend prototype.
• Stack kills version-drift hallucinations — the moment you write "Node" without a number, Claude assumes latest and generates syntax you don't support.
• Forbidden is your rollback log turned into policy.
• Verify stops Claude inventing scripts that aren't in your package.json.
• Commit style saves ~2 hours/week reviewing git history when Claude batches changes.

The six lines that carry 80% of the weight

If you're not going to write the full file today, write these six. This is the minimum viable version I hand to clients on a Loom call when we don't have an hour:

# Project
This is a [Node/Python/Go] service in production. Real users, real money.

# Stack
Node 20, Postgres 15, Prisma 5.7, Fastify 4, TypeScript strict.

# Forbidden
- Never run migrations, never install deps, never touch /billing or /auth without asking.

# Verify
Run: npm run test:unit && npm run lint && npm run typecheck.

# Commits
Conventional Commits + [TICKET-ID] or [no-ticket] suffix.

# Ask-before rule
If a change touches >3 files or any forbidden path, stop and ask before writing code.

The ask-before rule is the one most people skip. It's the single line that turns Claude from a caffeinated intern into a senior engineer. Without it, Claude finishes tasks. With it, Claude checks in when the blast radius is unclear — which is exactly what a good contractor does.

Live demo: same repo, same prompt, file on vs file off

I ran this test on a fresh clone of the invoicing repo. Prompt: "Add a discount_percent field to invoices."

File deleted, no CLAUDE.md:

• Claude opened the DB, decided to write a raw SQL migration
• Proposed /db/migrations/2026_add_discount.sql — a folder that doesn't exist
• Imported joi for validation — we use zod
• Added the field but skipped the TypeScript type update
• Two hallucinations, one silent bug, 30 seconds elapsed

Same clone, CLAUDE.md restored:

• Claude paused ~2 seconds reading the file
• Opened /prisma/schema.prisma (correct file)
• Proposed a schema change, not a migration file
• Stopped before running prisma migrate dev and asked for approval
• Added the zod schema in the right validator module
• Ran npm run typecheck when done — the exact command from section 4

Same model. Same prompt. Same repo state. The only variable was 80 lines of markdown. That's the entire lesson.

Comparison: course setup vs CLAUDE.md setup

I mapped what the top three Claude Code courses cover against what actually determines shipping quality. Numbers below are from my own client projects over 90 days:

The ROI is not close. The courses aren't wrong — they're just optimized for people who haven't shipped yet. The moment you have a live client, the file is the whole game.

Common mistakes when writing your first CLAUDE.md

The first version most people write is too long, too vague, or written like a pitch deck. Three patterns I see on every client audit:

• Marketing prose in section one. "Our mission is to empower small businesses…" Claude doesn't need mission statements. It needs stakes. Write "production, revenue-generating, live customers" and move on.
• Forbidden list written in the abstract. "Be careful with migrations" does nothing. "Never run prisma migrate deploy" is a rule Claude can actually enforce. Every forbidden line must be a shell command or a file path glob.
• Missing the ask-before rule. Without a stop condition, Claude will finish. Finishing on a 15-file refactor at 11pm on a Friday is how prod goes down.

One more anti-pattern: don't put secrets, API keys, or client-identifying info in CLAUDE.md. It's committed to the repo. Treat it like a README, not a .env.

Why bizflowai.io helps with this

The CLAUDE.md pattern is the same shape we use when we build automation systems for small business clients at bizflowai.io — a tight instruction file per workflow that pins the stack, forbids the dangerous actions, and forces the agent to stop and ask before it touches anything with money or auth. It's the difference between an agent that ships client work for 90 days without incident and one that quietly rewrites your billing logic on a Tuesday afternoon.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.