Go beyond CLAUDE.md with skills, helpers, memory, and letting Claude run your routines.

How to Let Claude Run Routines Without a Messy Desk

In the second part of a two-part guide, Claude Code users learn how to move beyond CLAUDE.md by matching each instruction to the right file. The core idea is timing: folder notes apply only when Claude works in one folder, skills open only when a request matches, subagents do noisy research on their own desk, progress notes and memory carry work into the next chat, and hooks enforce rules that should not depend on interpretation. For readers, this is a practical way to hand over everyday work without flooding Claude's limited working space with instructions that only matter sometimes.

Pick the file by timing

Start at the top and stop at the first yes. If a rule matters everywhere, CLAUDE.md is the right place. If it matters only in one area, a folder note is better. A CLAUDE.md inside a folder is read when Claude works there, like a sign on one door rather than a poster in every room. That is useful for local limits, such as newsletters being under 500 words.

The next category is a skill: a written routine for one job, stored in a folder with a file called SKILL.md. The clever design is that Claude keeps only the list of skill names on its desk. When your request matches one, it opens the full steps. That means dozens of skills can exist without costing much attention until one is actually used.

The description line matters most because it helps Claude decide whether the recipe fits. A simple weekly update skill might say to read every file in notes/ from the last 7 days, group items into Done, Doing, and Stuck, keep the result under 200 words, use a friendly tone, avoid jargon, and check that every item came from the notes. It should also say not to invent anything. After a task goes well, a user can ask Claude to turn what they just did into a skill.

Slash commands are similar routines, but they only run when typed, such as /weekly-report. They are a safer choice for actions that should never start by accident, including sending, publishing, or deleting.

Keep messy jobs off your desk

Some jobs create a large amount of temporary clutter. If you ask Claude to read 30 web pages and summarize what they say about pricing, the raw research can fill the working space with details you do not need. A subagent avoids that. It is a helper with its own desk. It does the messy reading there and returns a short summary. The mess stays with the helper, while the main chat receives the one-page brief.

The pattern is also useful for fresh-eyes reviews and fact-checking. Instead of asking the same thread to reread everything, send a helper to inspect the work and report only what changed. That keeps the main conversation focused on the decision, not the noise.

Carry work into tomorrow

Every new chat starts with a clean desk. That is good for privacy and simplicity, but it is a problem for long jobs. For work that spans days, ask Claude to keep a progress.md as it goes. The file can record what is done, what is next, and what did not work. Tomorrow's chat can read that note and pick up where the previous one left off.

Claude Code also keeps a small automatic memory of things it learns about a user, such as a preference for short titles. The command /memory lets a user see and edit that record. This makes memory different from a normal note: it is not for the current task only, but for repeated preferences that should follow the user across chats.

Give Claude a way to check itself

For anything bigger than a quick fix, use a four-step routine: look, plan, do, check. First, ask Claude to look without changing anything. Then ask for a plan. Read the plan, because fixing a plan usually costs a sentence, while fixing finished work can cost an afternoon. After that, let Claude do the work and check the result.

The most useful habit is to give Claude a way to check its own work. That could be a word limit, a checklist, or an example to match. With a check, Claude can keep going until the result is right. Without one, the user becomes the checker.

Notes are advice. Claude follows them almost always, but almost is not enough for rules that must hold every time. For those cases, use a hook. A hook is a small rule the app enforces, regardless of what Claude thinks. If one slip would really hurt, make it a lock. Hooks live in Claude Code settings, not in an .md file. A user can describe the rule, such as blocking any edits to a published folder, and Claude can help set it up.

Common mistakes

1. Cramming every rule into CLAUDE.md. If a note only matters in one folder, use a folder note instead. If it only matters for one routine, use a skill.

1. Treating notes as guarantees for critical rules. Notes are usually followed, but for rules that must never break, use a hook.

1. Asking for action before looking and planning. For anything larger than a quick fix, have Claude inspect first, plan second, then change files.

1. Forgetting to give Claude a check. Without a word limit, checklist, or example, Claude cannot verify its own output, so the user has to review everything.

1. Leaving passwords in notes. Notes are not a secure vault. Keep secrets out of the files Claude reads.

1. Trusting too fast. Do the task together first. The third time you do the same task, save it as a skill. Then let it run alone, but only produce drafts at first. Claude Code can run on a schedule, with nobody at the keyboard, and later you can let it act. Keep it asking before anything that cannot be undone.

1. Fixing only the wrong answer. When Claude gets something wrong, update the note, skill, or check that caused the mistake. Each fix makes future chats better.