Author(s): Somnath Mahanta

Originally published on Towards AI.

Beyond CLAUDE.md: Skills, Helpers, Memory, and Letting Claude Run Your Routines

In Part 1 we met CLAUDE.md, the note Claude reads at the start of every chat, and Claude's desk: the limited space where everything it's working with has to fit.

So where do you put everything else? Your step-by-step weekly report. Your long brand guide. What you were halfway through yesterday. Cram it all into CLAUDE.md and the desk gets messy again.

The answer: Claude has a few more .md files, and each one lands on the desk at a different moment.

The whole map

Pick the file by asking one question: when does Claude need to know this?

Folder notes: a sign on one door

Some rules only matter in one place: “newsletters are under 500 words.” Put a CLAUDE.md inside that folder, and Claude only reads it when it works there. A sign on one door instead of a poster in every room.

Skills: Claude’s recipe book

A skill is a written routine for one job, like “write my weekly update”. It lives in a folder with a file called SKILL.md.

The clever part: Claude keeps only the list of recipe names on its desk. When your request matches one, it opens that recipe and reads the full steps.

So you can have dozens of skills, and they cost almost nothing until one is used. A simple one:

---
description: Write my weekly update from this week's notes.
 Use when I ask for a weekly update or status report.
---
1. Read every file in notes/ from the last 7 days.
2. Group items into Done, Doing, and Stuck.
3. Under 200 words, friendly, no jargon.
4. Check every item came from my notes. Don't invent any.

The description matters most: it’s how Claude decides whether the recipe fits. And you don’t have to write skills yourself. After a task goes well, just say: “Turn what we just did into a skill.”

Slash commands are the same recipes, but they only run when you type them, like /weekly-report. Use them for anything you never want started by surprise: sending, publishing, deleting.

Subagents: send a helper to the library

Some jobs make a huge mess: “read these 30 web pages and summarise what they say about pricing.” A subagent is a helper with its own desk. It does the messy reading there and brings back a short summary.

Just ask: “use a subagent to research this.” They’re also great for fresh-eyes reviews and fact-checking.

Memory: a diary for tomorrow

Every new chat starts with a clean desk. For long jobs, ask Claude to “keep a progress.md as you go”: what’s done, what’s next, what didn’t work. Tomorrow’s chat reads it and picks up where it left off.

Claude Code also keeps its own small auto memory of things it learns about you, like “prefers short titles”. Type /memory to see and edit it.

Cheat sheet

Letting Claude run your routines

Look, plan, do, check

For anything bigger than a quick fix: ask Claude to look first without changing anything, then plan. Read the plan, because fixing a plan costs a sentence and fixing finished work costs an afternoon. Then let it do the work and check it.

The single most useful habit: give Claude a way to check its own work. A word limit, a checklist, an example to match. With a check, Claude keeps going until it’s right. Without one, you become the checker.

Signs vs. locks

Notes are advice. Claude follows them almost always, but “almost” isn’t enough for some things. For rules that must hold every time, use a hook: a small rule the app enforces, whatever Claude thinks.

Hooks live in Claude Code’s settings, not in an .md file. Just describe the rule ("block any edits to my published folder") and Claude can set it up.

Climb the ladder slowly

1. Do it together and correct as you go.
2. The third time you do the same task, save it as a skill.
3. Run it on its own, but only producing drafts at first. Claude Code can run on a schedule, with nobody at the keyboard.
4. Let it act, but keep it asking before anything you can’t undo.

Three safety habits

• Fresh desk per task. /clear starts a clean chat. Corrected Claude twice and it's still wrong? Start over with a clearer ask.
• Don’t hand it the keys. No passwords in notes. Be careful with web pages and emails it reads: strangers can hide instructions in them.
• When it gets something wrong, fix the note, not just the answer. Each fix makes every future chat better.

The short version

• Pick the file by when Claude needs the information.
• Skills are recipes; only their names sit on the desk.
• Subagents do messy jobs on their own desk.
• Progress notes and memory carry work into tomorrow.
• Notes are signs, hooks are locks.
• Look, plan, do, check, and give Claude a way to check itself.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI