I used to think the trick to getting more out of Claude was writing better prompts. Just keep talking, keep refining, keep pushing until it gets there. Turns out that’s the slowest way to work.
I came across a post from an AI professional who lays out a one-time setup that flips the whole thing around. Instead of prompting harder, the creator shows how to configure Claude once so it works like you do. I was genuinely impressed by how practical this is, so I broke down the whole system for you here.
Here’s the setup, point by point.
1. Switch from chat to Cowork
The original poster’s first move is to stop thinking of Claude as a single chatbox. Regular chat gives you one Claude and one reply. Cowork spins up several Claudes that plan, search, and build before they answer you.
Think of it as trading one assistant for a small team. To get there, the author says to download the Claude app, get Pro, and open Cowork.
2. Delete your old setup
This one surprised me. The expert admits folders and files were a mistake, his own included.
- Folders leak old work into new answers. Yesterday’s context bleeds into today’s task.
- Files rot the moment you stop updating them. Stale info quietly poisons your results.
His fix: turn OFF every setting and leave Global Instructions empty. Too much context makes Claude repeat itself. Give it room to think.
3. Build Skills from what you already do
Here’s where it clicked for me. The creator’s advice is to start with one task you already do every week.
- Do the task in a normal chat like you always would.
- Then prompt “make it a skill.” Claude builds it for you.
- Save it. Now you call it anytime with a /command.
Once you get the hang of it, the author suggests making more with /skill-creator and feeding them your real data. No coding, just teaching Claude your repeatable work once and reusing it forever.
4. Use Projects for context a Skill can’t hold
The next piece is Projects. A project holds the messy, specific context that a skill isn’t built for.
The industry pro says to drop in a client’s files, emails, and past campaigns. The project remembers all of it across every chat you have inside it.
The test the author gives is simple: if the thing is teachable to another person, make it a Skill. If it’s tied to a specific client or campaign, make it a Project. And yes, use your skills inside your projects.
5. Fix your habits, not just your tools
This was my favorite section. The mind behind this post is blunt: the setup is useless if you keep your old habits. A few of the shifts he recommends:
- Talk, don’t type. Install Wispr Flow, which is free, and speak your prompts, answers, and feedback.
- End every prompt with “ask me questions.” It forces Claude to close gaps instead of guessing.
- Call out the mistake. Don’t quietly accept a wrong answer, correct it on the spot.
- Export outputs to Google Drive. The author warns not to trust Claude’s own files as your storage.
Small changes, but together they turn Claude from a toy into a real work partner.
6. Save your Claude tokens
Long chats are token eaters, and the creator has a tidy set of rules to keep costs down and answers sharp:
- Don’t send follow-ups. Edit or restart from an earlier message instead.
- Start fresh often. Kick off a new session roughly every 20 messages.
- Batch your tasks. Bundle related asks into one prompt instead of drip-feeding.
- Match the model to the job. Use Sonnet or Haiku for quick tasks, and save Opus for the deep work.
That last tip alone can stretch your usage a long way without you noticing any drop in quality.
Why this setup matters
What I love about this breakdown is the core idea running through all of it. You’re not trying to out-prompt the model every single time. You’re setting it up once so it already knows how you work, what context matters, and which jobs you repeat.
The people who win with AI at work aren’t the ones typing the cleverest sentences. They’re the ones who built a system around it. This post is a clean roadmap for doing exactly that, and most of it takes an afternoon to set up.
If you want the full details straight from the person who shared it, check out the original LinkedIn post. It’s worth the read.