One email. No follow-up unless I ship something that fits.
Read firstThis is the Claude Code version. The source video names the files agents.md and memory.md. Claude Code reads CLAUDE.md and keeps its own memory folder, so the kit uses those names. The "four hours to one instruction" claim is the video's own. Test it on one real job.
THE FOUR FILES
What each file does and when it loads
Think of Claude as a new hire. Day one, you hand over four things. Here is where each lives and when Claude reads it, from Anthropic's Claude Code docs.
1. Instructions
CLAUDE.md in your project folder. Read at the start of every session. Keep it under 200 lines.
2. Context
A context/ folder of your offers, clients and prices. Claude opens a file only when the job needs it, if you point to it in plain words.
3. Memory
Notes Claude writes itself, in ~/.claude/projects/<project>/memory/. The first 200 lines or 25KB of its index load each session.
4. Skills
A .claude/skills/<name>/SKILL.md per job. Only a short description loads until you run the skill.
Know thisClaude treats CLAUDE.md as context, not a rule it must obey. For something that must always happen, Anthropic points to hooks instead.
SET IT UP IN TEN MINUTES
Six steps
Step 1Make a project folderAny folder you will run Claude Code in. Open a terminal there.
Step 2Add the four filesCopy the templates below into the exact paths shown. Fill every square bracket.
Step 3Start ClaudeType claude in that folder.
Step 4Check it loadedType /context and look for CLAUDE.md under Memory files. If it is missing, Claude cannot see it.
Step 5Try the skillType /weekly-client-report and give it one client's numbers.
Step 6Teach one more jobUse the worksheet near the end. One job, not ten.
your folderwhat you will end up with
CLAUDE.md context/offers-and-prices.md context/clients.md context/voice-samples.md context/tools.md .claude/skills/weekly-client-report/SKILL.md (memory lives outside the folder, see file 3)
FILE 1: INSTRUCTIONS
The file Claude reads first
Save as CLAUDE.md in the project root. It is written for an ads or agency business. Replace each square bracket.
CLAUDE filecopy and fill in
save as CLAUDE.md# [Agency name]: how Claude works here## Who we are [Agency name] runs paid ads for [type of client, for example local service businesses]. Team: [names and roles]. Timezone: [city]. ## How to write - Plain words. Short sentences. No hype. - Sign emails "[Your first name]". Never "Cheers". - Say what a number means, not just the number. ## How to work - If a request is unclear, ask one question at a time. - Show me a draft before anything is sent, posted or changed. - Never change a client's ad budget without my yes. - Say when you are guessing. ## Where things are Open these only when the job needs them: - context/offers-and-prices.md: what we sell and what it costs - context/clients.md: one section per client - context/voice-samples.md: three emails and two ads I wrote - context/tools.md: ad accounts, reports, who owns what ## Skills Jobs I taught once live in .claude/skills/. Run one with /skill-name. ## Rules learned When I correct you twice, add the rule here. - (none yet)
Plain pointers, not imports. A line starting with an at sign pulls that file in every time. A plain path in words does not, so Claude opens it only when needed.
Keep it short. Anthropic says files over 200 lines use more context and Claude follows them less.
Private notes go in CLAUDE.local.md, which you should add to .gitignore.
Run /init if you want Claude to draft a first file from an existing project, then trim it.
FILE 2: CONTEXT
Offers, clients, voice and tools
Four small files in a context/ folder. Fill the first two before you do anything else. The other two make the writing sound like you.
offers and prices filecopy and fill in
save as context/offers-and-prices.md# Offers and prices Last checked: [date] ## What we sell | Offer | Who it is for | What they get | Price | Minimum term | |---|---|---|---|---| | [Offer 1] | [client type] | [deliverables] | [price] | [term] | | [Offer 2] | [client type] | [deliverables] | [price] | [term] | ## What we do not sell - [Thing people ask for that you say no to] ## Common objections and my honest answer - "[Objection]": [what I say] ## Numbers I quote - Typical cost per lead for [client type]: [range] (source: [where it came from])
clients fileone section per client
save as context/clients.md# Clients## [Client name] - Business: [what they do] - Goal this quarter: [one line] - Monthly ad budget: [amount] - Contact and how they like to hear from us: [name, channel] - Never mention: [topic]
voice samples filepaste your real writing
save as context/voice-samples.md# Voice samples Paste real things I wrote. Claude copies the style, so use my best. ## Email to a new client [paste] ## Email when results are down [paste] ## An ad I am proud of [paste]
tools fileno passwords
save as context/tools.md# Tools and accounts - Ad accounts: [platform, account name, who has access] - Reporting: [tool, where the weekly numbers live] - Never put passwords here. Say where they are kept instead.
CatchAnything in your project folder may end up in git or be shared with a team. Keep client private data and passwords out of these files.
FILE 3: MEMORY
Notes Claude keeps for itself
Auto memory is on by default in Claude Code. Claude saves notes from your corrections without being asked. It writes four kinds: user, feedback, project and reference.
Easiest start: tell Claude in a session, "Remember: never sign emails with Cheers." Then type /memory and open the auto memory folder to see the file it wrote.
Or seed it yourself. Save these two files in that folder. The folder name comes from your repository, so use /memory to find the real path first.
memory indexthe index, one line per memory
save in the memory folder as MEMORY.md - [Email sign off](feedback_email_signoff.md): never end emails with "Cheers"
memory noteone memory
save in the memory folder as feedback_email_signoff.md--- type: feedback --- Do not sign emails with "Cheers". Sign off with my first name. Why: I said so twice. It reads as casual to clients.
Look at it monthly. It is plain markdown. Edit or delete anything wrong.
Turn it off with autoMemoryEnabled: false in settings, or set CLAUDE_CODE_DISABLE_AUTO_MEMORY=1.
Memory is stored on your machine only. It is not shared across computers.
FILE 4: A SKILL
Teach a job once, rerun it forever
A skill is a job you teach once. Save this as .claude/skills/weekly-client-report/SKILL.md.
skill fileone example skill
save as .claude/skills/weekly-client-report/SKILL.md--- name: weekly-client-report description: Write the weekly results email for one ad client from numbers I paste. Use when I say "weekly report for [client]" or paste a spend, leads and cost per lead table. disable-model-invocation: true ---# Weekly client report 1. Ask which client, and for this week's spend, leads and cost per lead. 2. Open context/clients.md and read only that client's section. 3. Compare to last week if I paste it. Say what moved and one likely reason. 4. Write the email in my voice from context/voice-samples.md. 5. Keep it under 150 words. Lead with the result. End with one next step. Checks before showing me: - Every number matches what I pasted. - No promise about future results. - Signed with my first name. Show the draft. Do not send it.
Run it by typing /weekly-client-report. The disable-model-invocation line means it runs only when you ask.
Claude sees only the description until the skill runs. That text is cut at 1,536 characters in the skill list. Mine is 163.
Personal skills that work in every project go in ~/.claude/skills/<name>/SKILL.md.
THE FIRST JOB WORKSHEET
Turn one real job into a skill
Pick one job you do every week that takes 30 to 60 minutes. Answer these in a note first.
1The jobName it in five words. For example, weekly client report.
2The inputsWhat do you paste in or open to start?
3The stepsWrite them as you do them, in order.
4The outputWhat does done look like? Paste one good example.
5The checksWhat do you always check before it goes out?
6The mistakesWhat went wrong last time? Claude should avoid it.
Then paste this into Claude Code with your answers:
promptcopy and paste
Interview me about this job one question at a time. My notes are below. Use them first, and only ask about gaps. When you have enough, write .claude/skills/[name]/SKILL.md. Start the description with when to use it. Add numbered steps, a short checklist, and one example output. Show me the file before you save it. Do not run the job yet. [paste your six answers]
Run the new skill on one real job and fix what is off. That is the test of the video's four hours claim, on your own work.
KEEP IT HEALTHY
Four habits
Correct twice, then write the rule. Add it under Rules learned in CLAUDE.md.
Check what loaded. Type /context after you change files.
Audit monthly. Type /doctor prompt-audit to find conflicting or outdated instructions. It needs Claude Code 2.1.283 or later, and changes nothing until you say so.
Already use AGENTS.md? Claude reads it only when there is no CLAUDE.md. Put @AGENTS.md at the top of CLAUDE.md to share one file.
SOURCES AND DATES
Where these facts come from
Anthropic, Claude Code docs, How Claude remembers your project, code.claude.com/docs/en/memory, read 25 Oct 2026.
YouTube Short "How to 10x your Claude with 4 .md files", youtube.com/shorts/ovLAIhbk3ek. About 838,000 views and 36.4 times the median of the channel's last 30 Shorts, per the YouTube listing, rounded. Publish date not retrieved.
Templates are mine, for an ads or agency business. I checked the file formats and line counts by script on 25 Oct 2026. I could not run Claude Code, so I have not run these files inside it.
That is the whole thing
Fill the two context files, teach one job, and correct Claude twice before writing a rule. Small and real beats big and generic.
I'm Raj Singh. I've led ecommerce at two Fortune 100 companies, and at one of them I ran about $8M a year in ads for four years. I've worked with AI since 2019. Now my engineering team and I ship AI products, including debtops.ai. Each guide here is written from a real install, with the output dated.
