My exact Claude starter structure (the 4 layers)

The aesthetic desk is the easy part. The thing that does the work is four files sitting in one folder, each doing a different job: one holds your context, one holds your playbooks, one holds your workflows in plain English, and one holds the tiny scripts that handle the boring parts.

The reason to build all four instead of one is that they feed each other. Your context file makes every skill sharper. Your directives stop you re-explaining a workflow. Your executions take the repetitive bits off your plate entirely. Start with layer one today and add the rest as you feel the need.

The four layers at a glance

On disk it is this small. One folder for the business, four things inside it.

my-business/
├── CLAUDE.md              ← layer 1: your context, always loaded
├── .claude/
│   └── skills/
│       └── weekly-report/
│           └── SKILL.md   ← layer 2: one playbook per job
├── directives/
│   └── weekly_report.md   ← layer 3: workflows in plain English
└── executions/
    └── pull_numbers.py    ← layer 4: tiny scripts Claude calls

CLAUDE.md, the context file

This is the layer everyone skips and it is the one carrying the other three. It is a plain document describing your work the way you would brief someone on their first day. Claude Code loads it automatically at the start of every session in that folder, which Anthropic's docs describe as "instructions you write to give Claude persistent context" (Claude Code memory docs).

Here is the whole file. Everything in square brackets is yours to replace, and the headings are worth keeping as they are.

# [Your name or business]

## What this is
[One or two sentences: what you do, who it is for, what you are building right now.]

## How I work with you
- Talk to me in plain language, infer what I mean, act, then summarise what you did.
- Ask before anything that deletes files, sends a message, or spends money.
- If you are unsure between two reasonable options, pick one and tell me why.

## My voice and rules
Voice in three words: [e.g. warm, direct, specific]
Always: [e.g. name real tools, never a vague category]
Never: [e.g. hype words, income claims, em dashes]
Two lines that sound exactly like me:
1. [paste a real line you wrote]
2. [paste another real line]

## The work I bring you most
1. [e.g. turning rough notes into a client update]
2. [e.g. drafting captions from a script]
3. [e.g. cleaning and sorting exports]

## Where things live
- Drafts: [folder or tool]
- Finished work: [folder or tool]
- Source of truth for [numbers, clients, content]: [where]

## Constraints
Tools I already pay for: [list them]
Budget ceiling: [number]
Claims I cannot make: [e.g. no income or medical claims]

If a blank template is the thing that stops you, do not write it yourself. Open any Claude chat and paste this instead, and it will interview you and hand the finished file back.

I want you to write the context file that every future chat with you will start from, and I would rather answer questions than stare at a blank template.

Interview me one question at a time and do not move on until my answer is specific enough to act on. Start with what I do and who it is for, then my voice and the rules I never break, then the three tasks I bring you most often, then where my files and numbers actually live, then my budget and the claims I am not allowed to make. If I give you a vague answer like "small business owners" or "I sound professional", push back once and ask me for the specific version with a real example.

When you have everything, write it up as one markdown document under the headings What this is, How I work with you, My voice and rules, The work I bring you most, Where things live, and Constraints, using my words rather than yours. Finish with a short list of anything you still do not know about my work, so I know exactly what to fill in later.

1. 📁 Make the folder. One folder for the business, anywhere you keep your work. Desktop is fine. This folder is the boundary of what Claude can see, which is the point.
2. 📄 Save the file as CLAUDE.md. At the root of that folder, named exactly CLAUDE.md. The name is not decorative, it is what makes Claude Code load it on its own.
3. ✅ Check it actually loaded. Start Claude Code in that folder and type /context. Look under Memory files for your CLAUDE.md. If it is missing, you are running from the wrong folder.
4. 🌱 Fix the file, not the output. Every time Claude gets something slightly wrong, add one line to CLAUDE.md instead of correcting it in the chat. That single habit is what makes the setup get better every week.

Skills, your reusable playbooks

A skill is a folder with one file inside it (a SKILL.md) holding a name, a one-line description of when to use it, and your instructions. Claude reads that description and pulls in the full skill the moment your request matches, so you teach a process once and never re-explain it (Agent Skills, official overview).

Start with the task you re-explain the most. Paste this and Claude drafts the whole thing, including the description line that decides when it triggers.

I want a reusable Claude skill for a task I repeat every week: [describe it, for example "turning my rough notes into our Monday client update"].

Write me the SKILL.md for it. Include a short lowercase-hyphenated name (do not put "claude" or "anthropic" in the name, or the upload gets rejected), a one-line description saying exactly what it does AND when you should use it, because that description is what makes you trigger it on your own, then clear step-by-step instructions in my style, then the output format so the finished thing looks the same every time.

Before you write it, ask me the three or four things you need to know about my format, my tone, and what I will hand you as input. When it is done, package it as a zipped folder I can upload, and tell me one sentence I could say in a chat that should trigger it.

This is the shape of what comes back, so you know what you are getting.

---
name: weekly-client-update
description: Writes our Monday client update. Use when I paste raw notes from the week and ask for the client update or the Monday email.
---
# Weekly client update

## Inputs
My raw notes for the week. If wins, blockers, or next steps are missing, ask me for them before writing.

## Steps
1. Pull the wins, blockers, and next steps out of my notes.
2. Write it in our format: one-line summary, then those three sections.
3. Put anything needing a decision from them in a separate short list at the bottom.

## Rules
- Plain and factual, no filler, no em dashes.
- Never invent a number. If a figure is missing, list it under "needs confirmation".

## Output
Show me the finished update inline, under 300 words.

1. 📦 Take the folder Claude hands you. A skill is that folder with the SKILL.md inside it. Ask Claude to zip it if it has not already.
2. ➕ Add it in the Claude app. Open Customize, then Skills, then Add, then Upload a skill, and pick the zip. Skills are available on Free, Pro, Max, Team and Enterprise plans (Anthropic's help centre), so there is no paywall to clear here.
3. 👩‍💻 Or drop it into Claude Code. Put the folder in .claude/skills/ inside your business folder for a skill only that project sees, or in ~/.claude/skills/ for one you want everywhere.
4. 🔁 Run it, then tune it. Use it on a real task, read what comes back, and tell Claude the one thing to change. Two or three passes and it does the job your way permanently.

Directives, your workflows in plain English

A directive is a markdown file describing one whole workflow in the order it actually happens, written the way you would explain it to a person. Mine live in a directives/ folder, one file per workflow, and I point Claude at the file instead of narrating fourteen steps again.

The difference from a skill is worth getting right. A skill is a job Claude runs on command and triggers itself. A directive is the operating manual for a process that involves you, your tools, and sometimes several skills in sequence. Skills are the verbs, directives are the recipes.

# [Workflow name, for example: turn a filmed video into a posted reel]

## What this is for
[One or two sentences: when I run this and what "done" looks like.]

## Before you start
- [What must already exist, e.g. the raw file is in /raw]
- [What to ask me for if it is missing]

## The steps
1. [First step in plain English, with the real rule, e.g. "cut every take where I stumble, keep the last clean one"]
2. [Second step, naming the tool or the script it calls]
3. [Third step, with the decision I want made and how to make it]

## Rules that must hold
- [The non-negotiables, e.g. never publish without showing me the final version]
- [The format, e.g. sentence case, no em dashes]

## What you hand back
[Exactly what I should see at the end: a file, a summary, a list of decisions I need to make.]

The fastest way to write your first one is to describe the workflow out loud and let Claude structure it.

I am going to describe a workflow I do regularly, in the messy order it lives in my head, and I want you to turn it into a directive file I can reuse.

Here is the workflow: [describe it, rambling is fine, include the tools you open and the decisions you make along the way].

Ask me about every step where you are guessing, especially the decision points where I choose between two options, because those are the parts I re-explain every time. Then write it as one markdown file under the headings What this is for, Before you start, The steps, Rules that must hold, and What you hand back. Write the steps in plain English rather than code, name the real tool at each step, and flag any step that is repetitive enough that it should become a small script instead of instructions.

Executions, the tiny tools

An execution is a small script that does exactly one thing and hands back a result. Transcribe a file. Rename a folder of exports. Pull this month's numbers into one table. Each one is single-purpose on purpose, so a directive can call three of them in a row and you can swap one without breaking the others.

You do not write these. You describe the one boring job, Claude writes the script, and from then on the directive calls it by name.

I want a small single-purpose script for one repetitive job, and I want it simple enough that I can read it and understand what it does.

The job: [describe it exactly, for example "take every .mp4 in a folder called raw, transcribe it, and save one .txt per video into a folder called transcripts"].

Write it in Python, make it do this one thing only, and have it print a plain-English line for every file it touches so I can see what happened. Handle the obvious failure cases out loud rather than crashing: a missing folder, a file that is already done, a file it cannot read. Tell me exactly what to install first and the one command I type to run it, and explain in two sentences what the script does so I could describe it to someone else.

Before you write anything, ask me where the files live and what I want the output named.

1. 📁 Make an executions folder. One folder, all your tiny tools in it, one file per job. Name them for what they do: transcribe_videos.py, not script2.py.
2. ▶️ Run it once by hand. Ask Claude for the exact command and run it on a copy of your files first. Watch the printed lines to confirm it did what you expected.
3. 🔗 Point the directive at it. Go back to your directive and replace that step with "run executions/transcribe_videos.py". Now the workflow file names the tool instead of describing the work.
4. 🌱 Only build one after the second time. The rule I use: if I have done a task twice by hand, it earns a script. Before that, the setup time is not paid back.

Two homes for the same structure

Same four layers, two places to put them. Start wherever you already are and move when you feel the ceiling, which usually arrives the first time you want Claude to touch a real file instead of talking about it.

The one honest hour

This whole structure started with one hour of writing down how I already worked. Not designing a system, describing one. Here is that hour, in order.

1. 🧠 Minutes 0 to 20, the context file. Run the interview prompt from layer one and answer honestly. Save it as CLAUDE.md. Incomplete beats absent by a mile.
2. 🛠️ Minutes 20 to 40, one skill. Pick the single task you re-explain the most, run the skill prompt, install it, and test it on something real.
3. 📝 Minutes 40 to 55, one directive. Describe your most annoying multi-step workflow out loud into the directive prompt. Note which steps it flags as script-worthy.
4. 🌱 Minutes 55 to 60, pick your next one. Write down the one job you will turn into an execution the next time you catch yourself doing it by hand. That is layer four queued up, no pressure.

Proof

Claude Code memory docs: what CLAUDE.md is and how it loads
Agent Skills, official overview
Creating custom skills, and the plans they are on
Claude Projects, including the free-plan cap
Claude Code setup, install commands and account requirements