Lesson completed!
-

Lesson 5 / 21 · Setup

The CLAUDE.md Brain

Max Techera
Next

The CLAUDE.md Brain

Last lesson you ran /init and it created a CLAUDE.md. This lesson is about what to put in it — because that file is the difference between an agent that knows your project and one that starts from zero every time you open it.

CLAUDE.md is the persistent brain. Claude reads it at the start of every session, before you type a word. It's where your niche, your voice, the accounts you track, and your rules live. Write it once, and every future session already understands your project.

A routing table, not a novel

The most common mistake is treating CLAUDE.md like documentation and dumping everything into it. Don't. Keep it under ~200 lines. The mental model is "a routing table, not a novel" — short, high-signal pointers that tell Claude where things are and how to behave, not long prose it has to wade through.

Warning:

A bloated CLAUDE.md makes the agent worse. Every line competes for the model's attention. If you wouldn't put it on an index card for a new teammate, it probably doesn't belong here. Link out to details; keep the brain lean.

Serious builders keep this tight on purpose. Doneyli De Jesus runs 15 agents and 24 skills entirely in Claude Code for about $100/month, and one of his hard rules is CLAUDE.md under 300 lines. Lean context is a feature, not a limitation.

The data folder structure

Your CLAUDE.md should point to where data lives. Here's the structure the whole system uses — four folders, each with one job:

  • data/mine/ → your reels (real, owner-only analytics pulled from your account)
  • data/refs/ → competitor posts, organized by creator
  • vault/hooks.json + ctas.json extracted from the outliers
  • dashboards/ → the generated HTML dashboards

The important design choice: data/mine/ and data/refs/ share one schema. That's what lets the vault rank hooks by pulling across both your content and your competitors' — instead of two disconnected piles of files. You don't have to build these folders now; you just declare them so Claude knows the plan.

The sample CLAUDE.md

Here's a complete starter you can adapt. Fill in the brackets with your niche and voice, list the accounts you want to track, and keep the rest as-is:

# Mi Content OS
Nicho: [tu nicho]. Idioma/voz: [rioplatense, directo, muestro receipts].
Competidores/referencias que trackeo: @cuenta1, @cuenta2, @cuenta3
Estructura de datos:
- data/mine/     → mis reels (analytics reales, owner-only)
- data/refs/     → posts de competidores (por creador)
- vault/         → hooks.json + ctas.json extraídos de los outliers
- dashboards/    → los HTML generados
Regla: nunca publicar nada automático. El sistema sugiere, yo decido.

Notice how little it is. Niche and voice in one line each, three accounts, the folder map, and one rule. That last line — "never publish anything automatically. The system suggests, I decide" — is the guardrail we'll come back to. It belongs in the brain because it's a rule Claude should honor in every session.

Tip:

Paste this in, swap the brackets for your real niche and accounts, and save. That's your brain. You can grow it later, but resist the urge — every session that starts clean is proof it's working.

Knowledge check

What's the right way to think about CLAUDE.md?

Key takeaway

CLAUDE.md is the persistent brain Claude reads every session — keep it under ~200 lines, a routing table not a novel. Point it at your four data folders (data/mine/, data/refs/, vault/, dashboards/), and bake in the one rule that matters: the system suggests, you decide. A lean brain is a better brain.

Share