AI Mayak
Start

Library · Memory and context: keeping the agent on track

The philosophy of folders: the PARA system for an AI entrepreneur

Confident user45 minUpdated: October 2026
18 of 105 in the library

Module: Second brain and knowledge management | Time: ~20 min theory + 25 min practice

The gist

Folder structure isn't an organizing question, it's an architecture question: Claude Code thinks about your project exactly the way your folders are laid out, and the right hierarchy multiplies the quality of your agents' work.

🎨 Picture this: Two offices. In the first, unlabeled boxes, documents piled together. In the second, clear departments, every cabinet labeled, every folder in its place. A new hire in the first office spends a day getting oriented. In the second, 15 minutes. Claude Code is that same new hire in every session.

Key concepts

  • PARA is a system of 4 categories that covers 100% of any information
  • Folder structure = the architecture of thinking: the way it's organized is the way it gets thought about
  • Claude Code reads the structure through CLAUDE.md, which is the agent's map
  • Golden rules for naming files make AI search more accurate
  • Flat vs. deep hierarchy: each approach has its own use cases

Theory

The PARA system: four boxes for everything

PARA was developed by Tiago Forte and has become a standard for the "second brain." The idea: every piece of information goes into one of four boxes.

Letter Name What goes there Example
P Projects Active projects with a deadline Launching a new product by June 1
A Areas Areas of responsibility with no deadline Health, finances, marketing
R Resources Reference material Guides, templates, research
A Archive Finished and no longer relevant Old projects, drafts

🎨 Picture this: P is a construction site with a completion date. A is your utility bills (you pay every month, there's no deadline). R is the reference books on the shelf. A is the storage unit with the old furniture. Everything you have goes into one of these four places.

The key principle of PARA: information is organized not by topic (like school notebooks) but by how active it is. This matters for AI: the agent understands that what's in Projects is alive and important, and what's in Archive is dead and optional.

Why folder structure = the architecture of thinking

Folder structure affects how you (and your agents) make decisions.

A bad structure:

Code
/projects/
  /my_business/
    idea1.txt
    old_idea.txt
    notes_final.docx
    notes_final_v2.docx
    notes_REALLY_final.docx
    random_stuff/
      ...

Claude in this structure: "I can't tell what's current, what's outdated, what's important. I'll have to guess."

A good structure (PARA):

Code
/projects/
  /launch-product-x/        ← active project
    BRIEF.md                ← context for Claude
    /research/
    /drafts/
    /final/
/areas/
  /marketing/               ← ongoing area
  /finance/
/resources/
  /templates/               ← reusable
  /guides/
/archive/
  /2025-product-y/          ← finished

Claude in this structure: "I can see active projects, reference material and an archive. I'll work in /projects/launch-product-x/ and use templates from /resources/templates/."

How Claude Code understands folder structure through CLAUDE.md

When it starts, Claude Code reads CLAUDE.md in the project root. It's the agent's instructions: what is where, what things are called, what the rules are. More on the file itself: CLAUDE.md: your project's system prompt. You can create a starter version with the /init command. Besides CLAUDE.md, Claude Code reads AGENTS.md and the rules in the .claude/rules/ folder, and keeps its own memory (auto memory), which you can manage with the /memory command.

An example CLAUDE.md for a PARA structure:

Type this into the chat
# Project Context

## Structure
- `/projects/` — active projects (each in its own folder with a BRIEF.md)
- `/areas/` — ongoing areas of responsibility
- `/resources/` — templates, guides, reference material
- `/archive/` — finished projects (do not edit)

## Naming Convention
- Folders: kebab-case, always lowercase (product-launch, not ProductLaunch)
- Files: kebab-case + a date if versioned (report-2026-10.md)
- UPPERCASE: only CLAUDE.md, README.md, BRIEF.md (important instructions)

## Rules
- Never edit `/archive/`
- Save drafts in the `/drafts/` folder inside the project
- Final versions go in `/final/` with no _v2 or _final suffixes

With a CLAUDE.md like this, the agent works precisely: it knows where to look, where to save and what not to touch.

Golden rules for naming files for AI

AI agents are like search engines: they work better with predictable names.

Rule 1: kebab-case everywhere

Type this into the chat
✅ market-research-2026.md
❌ Market Research 2026.md
❌ marketResearch2026.md
❌ market_research_2026.md

Rule 2: The date goes first for time-bound files

Type this into the chat
✅ 2026-10-01-competitor-analysis.md
❌ competitor-analysis-may.md

Dates first = automatic chronological sorting.

Rule 3: A verb or a noun, not "final"

Type this into the chat
✅ landing-page-copy.md
✅ email-sequence-onboarding.md
❌ landing_FINAL_v3_use_this.md
❌ email_copy_new2.docx

Rule 4: UPPERCASE only for the main instructions

Code
CLAUDE.md     ← the agent reads it first
README.md     ← a person reads it first
BRIEF.md      ← project context

Rule 5: No spaces in folder names

Code
✅ /my-business/
❌ /My Business/  ← breaks bash commands, confuses agents

An example structure for an AI entrepreneur

Code
~/workspace/
├── CLAUDE.md                    ← main instructions
├── projects/
│   ├── saas-tool-launch/
│   │   ├── BRIEF.md             ← context for Claude
│   │   ├── research/
│   │   │   └── 2026-09-competitors.md
│   │   ├── drafts/
│   │   │   └── landing-page-v1.md
│   │   └── final/
│   │       └── landing-page.md
│   └── youtube-channel/
│       ├── BRIEF.md
│       └── content-calendar.md
├── areas/
│   ├── marketing/
│   │   ├── brand-voice.md       ← Claude reads it before writing
│   │   └── target-audience.md
│   ├── finance/
│   │   └── budget-2026.md
│   └── tech-stack/
│       └── tools-and-keys.md
├── resources/
│   ├── templates/
│   │   ├── blog-post.md
│   │   ├── email-sequence.md
│   │   └── project-brief.md
│   └── guides/
│       └── claude-code-cheatsheet.md
└── archive/
    └── 2025-old-project/

🎨 Picture this: a well-organized kitchen. The cook (Claude) knows: spices are on the left shelf, fresh ingredients are in the fridge, the recipe archive is in the drawer. He doesn't search, he cooks. In a bad kitchen, the cook spends 30% of his time looking for the right tool.

Flat vs. deep hierarchy: which is better for AI systems

A flat structure:

Code
/projects/
  project-a-research.md
  project-a-draft.md
  project-b-research.md

✅ Good for: small projects, quick search, up to 50 files ❌ Bad for: big systems, collaboration, scaling

A deep structure:

Code
/projects/
  /project-a/
    /phase-1/
      /research/
        /primary/
          /interviews/

✅ Good for: big projects with many stages ❌ Bad for: everyday work (too many steps); it's harder for the agent to find its way at a depth of 4+ levels

The rule: at most 3 levels of nesting

Code
/root/
  /category/          ← level 1
    /project/         ← level 2
      /file.md        ← level 3

Three levels are enough for any solo project. Go deeper only for enterprises with teams.

Practice

  1. Take an honest look at your current structure:
bash
find ~/workspace -maxdepth 3 -type d | head -30
# or, if you don't have a workspace:
ls -la ~/Desktop
  1. Create a PARA structure for your main project:
bash
mkdir -p ~/workspace/{projects,areas,resources,archive}
mkdir -p ~/workspace/projects/my-project/{research,drafts,final}
mkdir -p ~/workspace/areas/{marketing,finance}
mkdir -p ~/workspace/resources/{templates,guides}

  1. Create a CLAUDE.md in the root of ~/workspace/ (take the example from this lesson and adapt it to yourself)

  2. Create a BRIEF.md in the folder of your active project:

markdown
# Project Brief: [Name]

## Goal
[What we're building and why]

## Audience
[Who it's for]

## Deadline
[When]

## Context for Claude
[What Claude needs to know before starting work]

## Constraints
[What's off-limits, style, budget]
  1. Start Claude Code in the ~/workspace/ folder and check that it understands the structure correctly:
Type this into the chat
Read CLAUDE.md and explain to me: if I ask you to write a draft
landing page for project X, where will you save it and why?

Tools and resources

Key takeaways

PARA isn't a system for storing files, it's a system for storing attention: Projects need action now, Areas need regular attention, Resources get used as needed, Archive is for history.

Claude Code works only as well as it understands the project structure. CLAUDE.md is the agent's map; without it, the agent is like a new hire on day one with no org chart.

Three levels of nesting and kebab-case naming: two rules that make working with AI agents noticeably more accurate for minimal effort.

What's next

→ AI copywriting with a prompt pipeline: writing in your own voice

The mark stays in this browser only and is never sent anywhere. My progress