So I’ve been watching people torch their weekly Claude token allowance in two days flat, and most of them have no clue why. Then I scrolled past this brilliant breakdown from a savvy AI professional on LinkedIn who laid out the exact 7 mistakes draining everyone’s Cowork budget, and honestly, it was the clearest fix-list I’ve seen all month.
The original poster didn’t just list problems. They gave the kill-shot for each one. I’m going to walk you through every single point because this is the kind of post you want to save, print, and tape to your monitor.
The 7-Point Quick Hit (Straight From the Author)
- Stop opening Chat when you mean Cowork.
- You need the Claude desktop app, not the browser.
- The $20/month plan works. Pay $100 only for heavy use.
- Open the Cowork tab (sitting between Chat and Code).
- Point it at a folder called “Claude Cowork.”
- Build out your folder structure properly.
- Set it up once and stop paying for re-prompts.
Now here’s where this contributor really earned their flowers. They went deeper on each mistake with the actual fix. Let me break it down.
1. Chat instead of Cowork
Why it kills you: Chat forgets everything the moment you close the tab. Cowork remembers your folder, files, and global instructions forever.
- Click the Cowork tab
- Point it at your folder
- Done. Persistent memory, every session.
2. No anti-AI-writing file means Claude writes like AI
This one hit me hard because I’ve been guilty of it. The expert recommends dropping an “anti-ai-style.md” file inside your ABOUT-ME folder.
- Ban words like delve, harness, leverage, and “it’s not X, it’s Y”
- No em dashes allowed
- Max 3 sentences per paragraph
Suddenly your output stops sounding like every other ChatGPT post on your feed.
3. Vague Global Instructions do nothing
The mind behind this post points out that most people set up global instructions once, type something fluffy, and wonder why Claude ignores their voice. Fix it:
- Go to Settings, then Cowork, then Edit Global Instructions
- Reference three files: about-me.md, anti-ai-writing.md, company.md
- Tell Cowork to save all deliverables in an OUTPUTS folder
Now every response respects your rules without you re-typing them.
4. Opus 4.7 for everything burns tokens
I see this one constantly. People default to the most powerful model for a quick lookup and wonder why their limit evaporates by Tuesday. This LinkedIn creator’s model-matching guide:
- Haiku 4.5 for simple lookups. Cheapest model, perfect for fast questions
- Sonnet 4.7 for drafts, edits, short answers. Your daily workhorse
- Opus 4.7 plus Thinking only for hard work. Strategy, deep analysis, code architecture
Match the model to the job, not the job to the model. That single shift can triple your weekly capacity.
5. No Skills means re-prompting 20 times for the same task
Skills are the unsung hero of Cowork, and the post’s author nails the setup:
- Create your first Claude Skill (the original poster mentions a free resource to build one)
- Set a trigger like “/linkedin”, “/contract”, or “/sales”
- One good Skill replaces an entire SOP document
I think this is the highest-leverage tip in the whole post. You build it once, you use it forever.
6. Zero plugins means Cowork is running blind
This contributor reminds us there are 11 official plugins just sitting there waiting:
- Type “/” to see all 11 available plugins
- Install only the 2 that match your actual job. Don’t install all 11
- MCP connectors plug into Gmail, Slack, and more. Free
Plugins give Cowork eyes and hands into your real workflow.
7. No scheduled tasks means you’re doing repetitive work by hand
This is the one that genuinely changed how I think about Cowork. The post’s author suggests treating Cowork like a junior employee with a calendar:
- Daily 7am: “Summarize yesterday’s emails”
- Monday 9am: “Pull last week’s LinkedIn metrics”
Monday starts with the report already drafted. That’s the magic. The work happens while you sleep.
Why I’m So Hyped About This Breakdown
Most “how to use Claude” posts are surface-level fluff. This one from the LinkedIn user is operational. Every point comes with a checkbox-style fix you can implement in the next 10 minutes.
I was blown away by how much waste I had baked into my own workflow. I was running Opus on lookup queries. I had zero scheduled tasks. My global instructions were two vague sentences. Fixing just the model-routing alone bought me back hours of weekly capacity.
If you’re paying for Claude and feeling like you’re constantly hitting limits, the problem isn’t the plan. The problem is the setup. This savvy professional just handed everyone a checklist to fix it.
Go check out the original post for the full visual layout. Worth a save and a share with anyone on your team who lives inside Claude.