Every question, answered on one page.
Ryan “RyMac” McKinney · USMC · updated 22 August 2026 · 80 answers
80 questions, 80 straight answers. Every one is lifted from the post that explains it in full, and every post is linked. If you came here from a search, your question is probably in this list.
Agentic Context Engineering Without The Agent Swarm
What is agentic context engineering?
Agentic context engineering is deciding what an AI agent can see at the moment it takes an action, rather than just answers a question. It matters more than ordinary context engineering because the agent acts on what it has, and a missing piece becomes a wrong action rather than a wrong sentence.
Do I need multiple agents to do agentic work?
No. Multiple agents help when the work is genuinely parallel and the pieces do not need to agree with each other, and when there is a team maintaining the setup. Most small operations are doing sequential work that is badly filed, where a single agent reading a well-organised folder outperforms a swarm.
AI Agent Memory: You Probably Have A Filing Problem
What is AI agent memory?
AI agent memory is the ability for an agent to carry information across separate sessions, so it does not start from zero each time. It is distinct from the context window, which is what the agent can hold within a single session.
Why does my AI agent forget things mid-conversation?
That is usually the context window filling up rather than a memory failure. When the window is full the oldest information drops out of view, and the model keeps answering confidently without it.
Do I need an AI agent memory tool?
Only if your information is already written down somewhere the agent can find it. If a brand new session cannot locate the state of a project in about 2 moves, the problem is that nothing was recorded, and a memory layer has nothing to remember.
What An AI Context Window Actually Is (It Is A Desk)
What is a context window in AI?
A context window is the total amount of information an AI model can hold in front of it at one time, including your instructions, the conversation so far, and any files it has read. It is measured in tokens. When the window is full, older information falls out of view and the model answers without it.
What happens when the context window is full?
The oldest information drops out of view. The model does not announce this and does not say it is missing something. The answers keep coming, with the same confidence, just without that piece. That is why it feels like the AI suddenly got worse partway through a session.
What is the difference between a context window and a token limit?
They describe the same constraint from different sides. The context window is the space available. The token limit is the unit that space is measured in. A 200,000 token context window means the model can hold roughly 200,000 tokens of combined input and output at once.
Should I just use the model with the largest context window?
A bigger window helps, but it does not solve the underlying problem. Research on long-context performance has shown that models get worse at using information as the amount in front of them grows, so filling a large window with everything you own can make results worse rather than better. Deciding what goes in front of the model beats making room for more of it.
Anthropic On Context Engineering, Read By A Non-Developer
What does Anthropic say about context engineering?
Anthropic's guidance centres on deliberately curating what an agent can see rather than filling the context window. The consistent finding is that model performance is limited by the quality and relevance of what is in front of it, and that adding more material can reduce accuracy rather than improve it.
Do I need to be a developer to apply it?
No. The technical guidance assumes you will implement curation in code, but the underlying principle is organisational. Deciding what the model sees, and keeping everything else findable rather than loaded, can be done with plain text files and folders.
I Was A Broke Marine On Dial-Up. The System Never Changed.
Do you need modern AI tools to run a folder system?
No. The same arrangement of plain text files in folders worked on a secondhand computer over a dial-up connection in 2006 and works with an AI agent today. What changed is that something else reads the files now.
Do you need to know how to code to build a business this way?
No. Ryan McKinney has never written a line of code and runs several businesses on text files in folders, directing an AI to do the work.
Why use plain text files instead of software?
Software goes out of business, changes its prices, or locks the export. A text file in a folder you own keeps working regardless of what any company does.
Claude Code Folder Structure: The 11 Line Version
What folder structure should I use with Claude Code?
One file that is always read first. One file describing the current job and the decisions already made. Then folders for the material: clients, offers, standard procedures, archive. The first file should be short and only point at the second.
Why should CLAUDE.md be short?
Because it is read every time. If it carries rules, examples and instructions it becomes another thing to process on every request. Keeping it to a routing role means the agent spends its attention on the material rather than the map.
Claude Code Rules Files: Where Your Rules Should Actually Live
Where do Claude Code rules go?
In the instruction file that is read first. Personal rules that are true everywhere go in the file in your home folder. Rules about one project go in that project's file. A separate rules folder adds a step without adding certainty.
Why does Claude ignore my rules?
Usually because the rule cannot be checked by looking. Write in a friendly tone is not a rule, it is a wish. Never use an em dash is a rule, because either it is there or it is not.
Should I make a .claude rules folder?
Rarely. A folder only helps once the rules are too many to read every time, and at that point the real problem is that there are too many rules. Cut them before you file them.
What Happens When The Claude Context Window Is Full
What happens when the Claude context window is full?
The oldest information drops out of what the model can see. Some tools truncate it, others compact the earlier session into a summary. Either way detail is lost, the model is not notified, and it continues answering with the same confidence using only what remains.
How do I know if I have hit the context limit?
The usual sign is the model contradicting a decision made earlier in the same session, or reintroducing an option you already rejected. It will not tell you directly, because it cannot see what it no longer holds.
Can I increase the context window?
You can choose a model with a larger window, but that does not remove the problem. Long-context research shows models get worse at using information as the volume in front of them grows, so filling a larger window can reduce accuracy. Deciding what goes in front of the model works better than making room for more.
Claude Directory Structure: Where Everything Actually Lives
Where is the Claude folder on my computer?
Claude Code keeps its own files in a hidden folder called .claude in your home directory. The dot at the front is what hides it. It holds your personal instruction file, your settings, and anything you have added such as skills.
What is the difference between the .claude folder and my project folder?
The .claude folder in your home directory is about you and applies everywhere. Your project folder is about one piece of work. Claude reads both, which is why personal preferences belong in the first and project facts belong in the second.
How should I structure my own folders for Claude?
One file that is always read first, one file describing the current job, and folders holding the material. Name folders the way a person would name them, and keep the count small enough that reading past them is cheap.
Can Claude read folders I do not tell it about?
It reads what it is pointed at and what it finds while working. That is why a map file at the top of the folder matters: it decides what gets read first rather than leaving the order to chance.
Claude Forgot What We Were Doing. Here Is Why.
Why does Claude forget what we were talking about?
Because the conversation is the only place that information lived, and a conversation has a size limit. When it fills, the oldest part stops being visible. Nothing was saved anywhere, so there is nothing to recall.
Will upgrading to Pro stop Claude forgetting?
Usually not. A paid plan gives you more usage, not permanent memory of a conversation. The forgetting is caused by information living only inside chat rather than in files the AI can re-read.
How do I stop Claude repeating itself?
Move the facts it keeps losing out of the conversation and into a file it reads at the start of every session. Once the answer is on the page, it stops being something the model has to remember.
CLAUDE.md Best Practices (From 137 Lines Of Getting It Wrong)
What are the best practices for a CLAUDE.md file?
Keep it short so it costs little to re-read. Make it point at the file holding current work rather than carrying that work itself. Write rules that are specific and testable. Put personal preferences in the home folder file and project facts in the project file. Delete anything that has not changed a decision in a month.
How long should a CLAUDE.md file be?
Short. Between 77 and 137 lines has covered every version of mine across several businesses. The file is re-read at the start of every session, so length is a cost you pay repeatedly rather than once.
Should rules go in CLAUDE.md or somewhere else?
Standing rules that never change go in the map. Anything about the current job goes in the desk file. Keeping both kinds in one file is the most common cause of an agent following the wrong instruction.
What Is A CLAUDE.md File? (And What Goes In It)
What is a CLAUDE.md file?
A plain text file, written in markdown, that Claude Code reads automatically at the start of a session. It carries the standing information about a project: what this is, where things live, and the rules you are tired of repeating. It is ordinary text and you can open it in any editor.
Where do you put a CLAUDE.md file?
At the top of the folder you work in. Claude Code also reads a personal one in your home folder that applies everywhere, and it reads a CLAUDE.md inside a subfolder when the work moves into that subfolder.
How long should a CLAUDE.md file be?
Short enough that it costs nothing to read every time. Mine has stayed between 77 and 137 lines across every version I have saved. Every time I made it longer the agent got slower to the point.
Does a CLAUDE.md file do anything on its own?
No. It does no work. Its only job is to be read first and point at the file that says what is happening right now. Making it do work is the most common way people break it.
CLAUDE.md Template And Real Examples
Is there a CLAUDE.md template I can copy?
Yes. The starting file is about 11 lines: what the folder is, an instruction to read the context file first, 3 standing rules in your own words, and the paths to where material lives. It is plain text and free.
What does a good CLAUDE.md file look like?
Short, pointing rather than instructing, and made of rules a person could follow without asking a question. A good one tells the reader where to go next in its first few lines.
Can I use the same CLAUDE.md for every project?
No. Preferences that are true everywhere belong in the personal file in your home folder. Each project's file should carry facts specific to that project.
Context Engineering Course: You Probably Do Not Need One
Do I need a context engineering course?
Probably not. Almost every course sold under that name teaches a developer to build the system in code. If you are the person who got handed the AI and cannot write code, what you need is a filing system, and that is two text files and a few folders.
What is context engineering in plain English?
Deciding what information your AI can see at the moment it does a job. That is the entire concept. Everything after that sentence is either implementation detail for developers or somebody selling a framework for a filing problem.
How long does it take to set up instead of taking a course?
About 20 minutes to build, and about a week of real work before you trust it. The structure is one file that is always read first, one file describing the current job, and folders holding the material.
Context Engineering vs Prompt Engineering: The Actual Difference
What is the difference between context engineering and prompt engineering?
Prompt engineering is how you phrase the request. Context engineering is what information the model can see while it answers. Prompting shapes the output. Context determines whether the model knows the facts it needs at all.
Is prompt engineering still worth learning?
Yes, for shaping output: format, length, what to omit, and what to do when uncertain. It is no longer the fix for a model that does not know your business, because no phrasing can supply information the model was never shown.
Context Engineering: Broken down barney style
What is context engineering?
Context engineering is deciding what information an AI model has in front of it at the moment it does a job. It covers what the model can see, what you choose to put there, and how it finds anything you did not put there directly.
What is the difference between context engineering and prompt engineering?
Prompting is what you say to the model. Context is what the model can see while you say it. Prompt engineering optimises the request. Context engineering optimises the material the request is answered against, which is now the bigger lever.
Is context engineering the same as RAG?
No. RAG, or retrieval augmented generation, is one technique for automatically putting relevant material in front of a model. It is a tool used inside context engineering, not a synonym for the practice.
Folder Structure For An Agency Running Multiple Brands
How should an agency structure folders for multiple brands?
Give every brand its own folder holding only the decisions that are different for that brand. Keep the rules that are the same for every brand in one shared folder, and have each brand point at it instead of copying it. Anything copied into 2 brands will drift apart.
How do I audit a folder structure?
Open a terminal in one brand folder and ask your AI a question only that brand can answer. If the answer is right, the structure holds. If it asks you a question the folder already answers, the routing is broken. Then check whether any fact appears in 2 files, because 2 copies of one fact is where drift starts.
What is the most common folder structure mistake?
Storing the same fact in 2 places. The moment a price, a colour, or a status lives in 2 files, one of them goes stale and nobody knows which. One fact, one home, and everything else points at it.
ICM Folder Structure: Every Folder, Named
What is the ICM folder structure?
2 plain text files and a small set of folders. One file is the map, read before anything else. One file is the desk, holding what is happening right now. The folders hold the material the work needs. It comes from Interpretable Context Methodology.
How many folders should I start with?
Four. Clients, offers, standard procedures, and archive. Add a fifth only when you have put something in it twice and it did not belong in any of the four.
Does the folder structure need to be on a particular drive or service?
No. It works anywhere you already keep files. What matters is that it is somewhere you look every day, because a filing system you have to remember to visit is not a filing system.
ICM Folder System Template: Take Mine
What is in the ICM folder system template?
6 files: a start page, the map file, the desk file, the folder skeleton, a 1 page setup sheet, and a list of what to cancel. It also contains a half built website project so you practise on real work rather than an empty folder.
Does the ICM folder template cost anything?
No. There is no card and no trial. It is a set of text files and folders, and charging for that would be the exact thing the site complains about.
Can I use the template with something other than Claude?
Yes. The template is plain text files and ordinary folders, so any AI that can read files can read it. Nothing in it is specific to one product.
What Is The ICM Folder System? (Files In Folders)
What is the ICM Folder System?
The ICM Folder System is Interpretable Context Methodology applied as plain text files in ordinary folders. One index file is always read first. One context file describes the current job. Folders hold the material. A single AI agent can then find what it needs without an orchestration platform.
What does ICM stand for?
Interpretable Context Methodology. It comes from a paper by Jake Van Clief and David McDermott on folder structure as agent architecture. A practical way to remember the parts is Index, Context, Material.
How long does it take to set up?
About 20 minutes. There is nothing to install. You place the folder where you already work, fill in four headings in one file, and point your AI agent at it.
ICM For People Who Cannot Write Code
Do I need to know how to code to use ICM?
No. The system is plain text files inside ordinary folders. Anything you can do in a text editor is enough.
Do I need to use a terminal?
No. You can create the folders and files the same way you would create any other document. A terminal is one way to work, not a requirement.
What is a markdown file?
A plain text file with a .md ending. A hash mark at the start of a line makes that line a heading. That is the entire format, and any text editor can open one.
The ICM Paper, In Plain English
What is the ICM paper?
Interpretable Context Methodology: Folder Structure as Agentic Architecture, by Jake Van Clief and David McDermott, submitted 17 March 2026. It argues that filesystem structure can replace the code most people write to coordinate AI agents. It is open source under the MIT license.
What is the Model Workspace Protocol?
MWP is the name the paper gives its method. Numbered folders stand for stages of work, plain markdown files carry the instructions and context for each stage, and ordinary scripts handle the mechanical steps that need no AI at all.
Does the paper say multi agent frameworks are bad?
No, and this is the part most summaries get wrong. It says those frameworks work well for complex concurrent systems, and that they are unnecessary overhead for sequential work where a human checks the output at each step.
ICM vs Prompt Engineering vs RAG: Which Is Your Actual Problem
What is the difference between ICM and prompt engineering?
Prompt engineering is about what you say to the AI. ICM is about what the AI can see while you say it. Better wording cannot fix missing information, and complete information makes wording matter less.
Is ICM an alternative to RAG?
They solve different problems. RAG fetches passages out of a large body of documents to answer questions about them. ICM arranges the working files so a single agent knows the state of the job. A business can honestly need both, but most need the second one first.
Which should I do first?
Whichever matches your symptom. If the AI has the facts and still answers badly, that is a prompting issue. If it keeps losing what it was told, that is an arrangement issue. If you need answers out of thousands of documents, that is a retrieval issue.
What Is Interpretable Context Methodology? (ICM, Plain English)
What does interpretable context methodology mean?
It means arranging the information an AI needs into plain files and folders that can be read and understood by a person as well as by the machine. Interpretable means you can open it and see why it says what it says. Context means the information the AI has in front of it while it works. Methodology means there is a repeatable order to it rather than a pile.
Is ICM the same as the ICM folder system?
ICM is the method. The ICM folder system is what the method looks like once it is set up on a real machine: one map file, one desk file, and folders holding the work.
Do I need to write code to use ICM?
No. It is text files and folders. The person writing this cannot write a line of code and runs several businesses on it.
Is The ICM Folder System Legit? The Case Against It First
Is the ICM folder system legit?
The method is real, published, and free under the MIT license. The honest criticism is not that it is fake, it is that the underlying idea of keeping organised files is not new. That criticism is correct and the paper itself says so.
Is the ICM folder system a scam?
No. The paper and the protocol are open source and cost nothing, and the folder structure can be built by hand in an evening without paying anybody. Paid groups that teach it are a separate purchase from the method itself.
How can I test it without spending money?
Set up 2 text files and 4 folders, run 1 real job through it, and compare against your current setup after a week. That test costs nothing and does not require trusting anyone's opinion.
Is Jake Van Clief Legit? A Paying Member Answers
Is Jake Van Clief legit?
Yes. He co-wrote the Interpretable Context Methodology paper with David McDermott, it is published on arXiv, and the method is open source on GitHub under a permissive licence. His community is free to join. I pay for the paid tier by choice and I have been a member since 1 July 2026.
Do you have to pay to learn the ICM folder system?
No. The paper is free on arXiv, the reference code is free on GitHub, and Jake's community has a free tier. Paying gets you his time and the room, not access to the method.
How can I check the ICM folder system myself?
Open a project folder that already has a map file and a desk file in it, point your AI at the folder, and ask it where the project stands. If it reads the files and reports back without you explaining anything, the structure is doing the work. That takes about 20 minutes and costs nothing.
Jake Van Clief's ICM Folder System, From A Guy Who Runs It
What is Jake Van Clief's ICM folder system?
It is Interpretable Context Methodology set up as plain text files in ordinary folders. One map file read first, one desk file holding the current job, and folders holding the material. Jake Van Clief and David McDermott published the underlying paper and Jake teaches the folder structure.
Do you have to join Jake's group to use it?
No. The structure is 2 text files and a handful of folders and you can build it yourself in an evening. The group is where the teaching and the room are, which is a different purchase from the method itself.
Is this the same as multi agent orchestration?
It is close to the opposite. It is one agent with a good map instead of several agents handing work between them.
The kit is 6 files. It takes 20 minutes. A start page. A map. A desk file. The folder skeleton. A setup sheet. A list of what to cancel. No card. No call. No course to buy.
Send me the folder system