Unspoken Rules That Make or Break Claude
Claude drifts not because the model is inconsistent, but because most projects give it nothing to hold. Five architectural layers stop the drift - this course walks through each one in full.
Session one is usually fine.
Claude reads the first message, understands the context, and produces output that matches what you were thinking. The model is sharp. The response lands. Something about it feels like the tool finally clicks.
Session three is where the pattern breaks.
The context has shifted. Decisions were made since the last session. Claude does not know what changed - what rules were established, what patterns were rejected, what standards are being held. It starts fresh. Because every session, it does.
This is not a model problem. The problem is that most projects give Claude nothing to hold between sessions. No structure. No rules. No memory. Just the prompt and whatever the model can infer from it.The drift that follows is predictable. Different choices on the same question. Patterns established last week, abandoned this week. Output that feels slightly off in ways that are hard to name. Most builders respond by writing longer prompts. The drift continues.
The fix is not a better prompt.
The five layers that stop the drift
There is a framework called the Agent Orchestration Layer - AOL for short. It maps how deeply Claude is integrated into a project across five levels. Each level adds a rule the agent follows automatically, without the builder repeating it.
Most builders operate at AOL 1 or AOL 2 and experience the result: an agent that requires constant correction, re-explanation, and oversight. The five layers are not features to add at random. They are a system. Each layer builds on the previous one. Skip a layer and the system does not compound - it leaks.
AOL 1 is the default state. A session opens, a message goes in, Claude responds. When the session ends, everything resets. This is how most people use Claude - which is why drift is the default experience.
AOL 2 introduces CLAUDE.md. This file sits in the project root and acts as the permanent rulebook. Because modern AI editors read CLAUDE.md before processing any instruction, Claude inherits the project’s standards from the first message of every session. Naming conventions, tech stack, forbidden patterns, tone rules - all of it carried forward automatically, without repeating a word.
AOL 3 extends memory beyond what a single file can hold. A real project accumulates complexity: decisions made months ago, bugs that cost days to diagnose, architectural choices with implications across dozens of files. A structured /docs folder gives Claude that depth - not as chat history, but as organized, agent-readable documentation that survives every reset.
AOL 4 introduces Skills. A Skill is a reusable instruction set stored in the project and invoked with a slash command. Instead of writing the same prompt every session to run a code review or generate a component, the logic is written once. The agent runs it identically every time, without variation, without interpretation.
AOL 5 adds Hooks - automated enforcement that runs before and after every tool call. A PreToolUse hook warns Claude before it touches a critical file. A PostToolUse hook validates the output and corrects a mistake before the file is saved. Hooks are the layer that turns a system you supervise into a system that supervises itself.
- Rules explained from scratch every session
- Patterns reset with every context window
- Mistakes repeated across sessions
- Agent requires constant supervision
- Rules read automatically from CLAUDE.md
- Decisions preserved in /docs across every reset
- Skills run repeatable tasks without re-prompting
- Hooks enforce standards without supervision
The fourteen lessons below walk through each layer in full - from CLAUDE.md to Hooks, plus the three frameworks that map where any Claude setup currently stands. Watch every lesson here, then download the free Starter Kit files to apply what you see.
The files require a free account. The videos do not.
Course contents
- Course Introduction - Why the course exists and what it builds
- Welcome to the Starter Kit - What is inside and where to start
- How to Activate Claude - The 3-step setup sequence
- Your AI Setup Roadmap - APL, AAL, and AOL explained
- CLAUDE.md - The project constitution
- The /docs Folder - Persistent project memory
- Advanced /docs - Scaling context at depth
- Claude Skills - Reusable instructions on command
- Claude Hooks - Automated rule enforcement
- MEMORY.md - Memory between sessions
- Setup Complete - Activating the system on a real task
- Agent Power Levels - APL framework deep dive
- Agent Access Levels - AAL framework deep dive
- Agent Orchestration Layers - AOL framework deep dive
Course Introduction
Claude starts clean every session. Fourteen lessons in this course build the structure that changes that - from zero context to a system that holds your rules, your memory, and your standards across every reset.
01 - Welcome to the Starter Kit
Ten files across three groups make up the Starter Kit - Frameworks, Starter Files, and Automation. This lesson maps what each group does and what order to follow. Download the Complete Bundle ZIP first, read the Setup Instructions second, then move to Lesson 2.
02 - How to Activate Claude
Three steps connect Claude to your project rules from the first message. The Master Setup Prompt, included in the Kit instructions, tells Claude to read your CLAUDE.md before processing anything else. After this lesson, the agent acknowledges your standards in the very first reply - not as a default response, but as confirmation that your rules are in place.
03 - Your AI Setup Roadmap
Three framework PDFs in the Starter Kit answer the question that information overload makes impossible: where does your setup actually stand, and what does the next level require. APL maps the type of setup. AAL maps the depth of access. AOL maps the integration layer. After this lesson, the frameworks replace guessing with a map.
The files for every layer in this course - CLAUDE.md template, MEMORY.md, SKILL.md, Hooks guide, /docs scaffolding, and the APL, AAL, and AOL framework PDFs. Free account required.
04 - CLAUDE.md
Without CLAUDE.md, Claude guesses the project standards every session and produces different patterns every time. One file in the project root changes that - the agent reads it before doing anything else and adopts the rules from the first message. This lesson covers what belongs in the file, how the template works, and what changes immediately after it is in place.
05 - The /docs Folder
CLAUDE.md is the entry point. The /docs folder is the memory. A single file cannot hold the architectural decisions made months ago, the bugs that took days to diagnose, or the constraints that apply only to specific parts of the project. This lesson explains when a CLAUDE.md alone is enough and when /docs becomes the difference between an agent that remembers and one that starts over.
06 - Advanced /docs
A flat /docs folder works for small projects. At scale, Claude needs context that is organized by intent - not just present. Numbered folders by domain, agent-ready files that answer specific questions, and a CLAUDE.md that points directly to /docs entries instead of repeating them. This lesson covers the patterns that make large documentation folders work without losing the agent’s focus.
07 - Claude Skills
Skills are reusable instruction sets that run with a single slash command. Write the logic once - for a code review, a component generator, a security check - and the agent runs it identically every time, without interpretation, without variation. The Starter Kit includes a SKILL.md template as the starting point for the first one.
The SKILL.md template, Hooks guide, MEMORY.md, and every other file covered in this course. Free account required to download.
08 - Claude Hooks
Skills define what Claude does. Hooks define what it cannot do - and enforce it automatically on every tool call without monitoring. A PreToolUse hook warns before critical files are touched. A PostToolUse hook validates the result and corrects a mistake before the file is saved. This is the layer that turns a system you supervise into a system that supervises itself.
09 - MEMORY.md
Without MEMORY.md, every session starts from zero - the same questions repeated, the same corrections made again. With it, Claude carries forward your preferences, your decisions, and your project context from the first message of every session. This lesson covers what belongs in the file - user context, feedback records, project decisions, references - and what does not.
10 - Setup Complete
Five layers are in place. This lesson covers the one step that determines whether the system compounds or stays static - running Claude on a real task for the first time, confirming the agent reads the rules, and adding the first MEMORY.md entry after the session. From here, each session adds memory and each correction strengthens the rules.
11 - Agent Power Levels
Four levels separate a browser chat session from a multi-agent orchestration system. APL (Agent Power Level) maps exactly which level any current setup is at - and what the next level actually costs in terms of risk, configuration, and control. After this lesson, the APL framework replaces tool comparison with a structural decision.
12 - Agent Access Levels
APL measures power. AAL (Agent Access Level) measures depth - how far the agent can reach into systems, files, and data. Five levels from classic chat with no access to self-healing infrastructure. The critical risk appears when access exceeds control: the agent can do more than the setup can safely manage. This lesson maps the five levels and the warning signs at each one.
13 - Agent Orchestration Layers
AOL is not about which model is running. It is about how integrated the agent is in the project - from stateless chat with no rules to automated enforcement that fires on every tool call. Each file in the Starter Kit corresponds to one AOL layer. This lesson closes the course by placing every file in the full framework map and showing the path from AOL 1 to AOL 5.
Every file covered in this course - CLAUDE.md template, MEMORY.md, SKILL.md, Hooks guide, /docs scaffolding, and the APL, AAL, and AOL framework PDFs. Free account required.