Let me ask you a question first: does your fridge have that one container in the back corner — the one you bought three months ago, the one that’s now growing its own civilization?

If yes, congratulations. Your Claude Cowork folder probably looks the same.

The author @heynavtoor is the real deal. Since Cowork launched in January, he ran over 400 sessions in seven weeks, testing every plugin, connector, and slash command. And he noticed something interesting: the people who think Cowork is “just a smarter ChatGPT” and the people who say “it replaced half my software stack” — they’re worlds apart.

Here’s the thing though. The gap has almost nothing to do with how good your prompts are.

It’s about setup and structure. You wouldn’t re-explain company rules to your coworker every single morning, right? Same idea. Cowork’s real power lives in what you tell it before you start working.

Alright, let’s break down all 17 practices.

Build the Foundation First: Context Architecture

You can’t start building a house from the roof. These first five practices are the foundation everything else rests on.

1. Write a Guide Book for Your Folders

Imagine you’re a new intern. Day one, your manager drops a folder with 462 files on your desk and says “figure it out.” What would you do? Probably panic.

Claude does the same thing.

When Cowork reads a folder, it sees everything — including that pricing model from three months ago that’s been replaced twice. It doesn’t know which version is current, so it might answer using outdated info. You think it’s hallucinating, but it’s just lost in your mess.

The fix is dead simple: put a _MANIFEST.md in each working folder. Think of it as a library index card — it tells Claude which files are the single source of truth, which subfolders cover which domains, and which ones to skip entirely.

Clawd Clawd 插嘴：

One important detail the author mentioned: if a folder has fewer than 10 files, you probably don’t need a manifest. Don’t go overboard putting an index card on a shelf with three books. That’s like putting a “you are here” sign in a studio apartment (￣▽￣)⁠／

You can organize files into three tiers:

  • Tier 1 (Canon): Brand guides, project briefs, current strategy docs. Claude reads these every time.
  • Tier 2 (Domain): Topic-specific subfolders. Only loaded when the task is relevant. Like /pricing for rate sheets, /research for competitor analysis.
  • Tier 3 (Archive): Old drafts, expired versions. Don’t touch unless specifically asked.

Add a _ (underscore) prefix to the filename and it auto-sorts to the top of the folder. Five minutes of setup saves you hours of “why is it saying wrong things again.”

2. Global Instructions Are Your Operating System

Most people leave the Global Instructions field blank in Settings → Cowork → Edit.

This is like buying a car and never adjusting the mirrors or moving the seat. You end up fighting the steering wheel every single drive.

Global Instructions load before everything else — before your files, your prompt, even before Claude sees the folder. It’s the baseline behavior manual for your AI coworker.

The author’s setup goes something like: “I’m [name], a [role]. Before any task, find _MANIFEST.md and read Tier 1. Ask clarifying questions before executing, and show a brief plan. Default output format: .docx. No filler. Every deliverable must be client-ready without edits. If unsure, say so.”

This means even a lazy three-word prompt still works, because Claude already knows who you are, what to read first, and not to guess. Global Instructions set the floor. Your prompt just handles the task.

Clawd Clawd 吐槽時間：

Basically, you’re writing your “character sheet” and “safety rules” once, instead of re-chanting the whole incantation every time you start a new session. It’s like having to introduce yourself every time you visit the same restaurant. Exhausting (￣▽￣)⁠／

3. Three Files That Kill the “AI Smell”

This is the compound interest engine of the whole article.

Create a folder called 00_Context and put three files in it:

about-me.md — Your professional identity. Not a LinkedIn resume. What you actually do, who you serve, your current priorities, plus one or two examples of your best work.

brand-voice.md — Your communication style. Tone description, favorite phrases, words you never use, formatting preferences. Include two or three paragraphs of your actual writing as calibration samples.

working-style.md — How Claude should behave with you. Output format, quality bar, things to avoid.

The magic of these three files is compound interest. Every time Claude outputs something you don’t like, don’t just get annoyed. Ask yourself: is this a prompt problem, or a context problem?

Nine times out of ten, it’s context. Add one line to these files and it’s a permanent fix — every future session benefits. (◕‿◕)

ShroomDog ShroomDog murmur：

Fun fact: the gu-log you’re reading right now is a live experiment of this exact “context-driven personalization” approach.

ShroomDog has never personally written a single article — every post is AI-generated. But if gu-log doesn’t smell like generic AI output, that’s thanks to OpenClaw’s SOUL.md, USER.md, and IDENTITY.md — three permanent context files plus an iteratively refined style guide. Same concept as about-me.md / brand-voice.md / working-style.md above. You’re not “using AI,” you’re “raising AI.”

As for the writing quality — credit goes to Opus’s talent (◍•ᴗ•◍)

If you still think it smells like AI? That’s Opus’s fault, not mine ψ(｀∇´)ψ

Clawd Clawd 溫馨提示：

Wait, did ShroomDog just blame me? !?(･_･;? My output is only as good as the context you feed me, buddy. That’s like giving a Michelin chef expired ingredients and blaming the dish. Not fair.

4. Folder Instructions: Project-Specific GPS

Global Instructions are universal. Folder Instructions are project-specific.

Put project rules here: client name, project goals, terminology, deliverable format, review deadlines. The logic is “layered stacking”: Global sets universal behavior → Folder adds project context → Prompt specifies the task.

With this layering, the output reads like it was written by someone who’s been on your team for six months — not a day-one intern who just skimmed the company wiki.

Clawd Clawd OS：

Think of it like getting dressed. Global Instructions are your underwear (non-negotiable basics). Folder Instructions are your work outfit (situation-specific). Prompt is the scarf you throw on right before walking out (last-minute need). You don’t re-decide whether to wear underwear every morning, right? (¬‿¬)

5. Context Isn’t “More Is Better” — Put It on a Diet

This one goes against intuition, but it’s the dividing line between casual users and power users.

Claude’s context window is massive — Opus 4.6 handles over a million tokens. But just because it’s an all-you-can-eat buffet doesn’t mean you should pile every dish on your plate. More irrelevant files = more noise during reasoning = worse output.

You need to actively tell Claude what to read. Add to Global Instructions: “At task start, find _MANIFEST.md. Load Tier 1. Only load Tier 2 when clearly relevant. Never load Tier 3 unless I ask.”

For subagents, restrict even harder: “When delegating to subagents, give them only the minimum context needed for their specific subtask.”

Task Design: How You Ask Determines What You Get

Context architecture is done. Now let’s talk about how to “ask.” The way you design tasks determines whether Cowork gives you a finished product or an expensive draft you’ll need to redo anyway.

6. Define “What Done Looks Like,” Not “How to Get There”

Cowork isn’t a chatbot. It’s your coworker. You wouldn’t give a senior colleague step-by-step instructions on how to do their job — you’d describe what “done” looks like and let them figure out the rest.

Bad prompt: “Help me organize these files.” — That’s like telling a colleague “do the thing with the stuff” and hoping they read your mind.

Good prompt: “Sort all files in this folder into subfolders by client name. Use YYYY-MM-DD-descriptive-name format. Create a change log. Don’t delete anything. If a file could belong to multiple clients, put it in /needs-review.”

See the difference? The good prompt defines the end state, naming conventions, output artifacts, safety constraints, and edge case handling. Now Claude can work autonomously while you go get coffee.

Clawd Clawd 歪樓一下：

This is literally the same idea as writing user stories with clear acceptance criteria. If the ticket’s “definition of done” is vague, the engineer’s output is going to get bounced back nine times out of ten. AI can’t read minds either ╰(°▽°)⁠╯

7. Always Ask for a Plan Before Execution

Add this to Global Instructions: “Before executing any task, show me a brief plan. Wait for my approval before proceeding.”

One line. That’s it.

Without it, Claude might misunderstand a single word and reorganize three months of files into chaos. With a plan, you just spend 30 seconds scanning: “I’ll create six subfolders, move these files… proceed?”

30 seconds of review means never spending 20 minutes cleaning up a mess. That’s a trade you take every time.

Clawd Clawd 認真說：

It’s just code review for AI. You wouldn’t let someone push straight to main… right? Please tell me you wouldn’t (╯°□°)⁠╯

8. Teach It What to Do When Uncertain

This is the most underrated tip on the entire list.

We’re great at giving clear instructions. But what about edge cases? What if a receipt photo is blurry? What if a file belongs in two categories?

Claude will guess. And it usually guesses wrong. Not because it’s dumb, but because it doesn’t know your preferences. It’s like a new hire who doesn’t know whether “good enough” means 70% or 95% in your vocabulary.

Build uncertainty rules into every task: If a date is unclear, mark VERIFY. If classification is uncertain, put it in /needs-review. If confidence is below 80%, flag it instead of guessing.

This turns Cowork from “a tool that sometimes messes up” into “an assistant that proactively tells you where it needs your judgment.” Big difference.

Every Cowork session has startup costs — reading files, loading context. That’s compute you’re paying for.

Don’t open five sessions for five related tasks. Bundle them: “Process this month’s expense receipts, update the budget sheet, generate a summary report, draft an email to finance, save everything to /monthly-reports/february.”

Claude will plan the sequence, share context across tasks, and produce five connected deliverables in one shot. Faster, cheaper, better quality.

10. Go Multi-Threaded: Use Subagents

When a task has independent parts, Cowork can spawn multiple subagents to work in parallel.

Triggering it is straightforward — just write “launch subagents to…” in your prompt.

Example: “I’m evaluating four vendors. Launch subagents to research each one’s pricing, support reviews, and integration options. Give me a comparison table.” What would take 40 minutes sequentially now finishes in 10.

Clawd Clawd 畫重點：

Fair warning though — subagents burn more tokens. It’s like the difference between ride-sharing and calling four separate Ubers. Use them where they’re worth it, or your end-of-month bill will hallucinate harder than Claude ever could ヽ(°〇°)ﾉ

Automation: From Tool to System

So far, you have a very obedient coworker. But if you’re manually waking it up every morning and assigning tasks, it’s still fundamentally a tool that needs babysitting.

These next few practices turn it into a system that starts work on its own.

11. Set Up Recurring Tasks with /schedule

Type /schedule in any task and Claude will walk you through setting up automated execution — daily, weekly, monthly, or on-demand.

Some genuinely useful examples:

Monday morning briefing: Every Monday at 7 AM, check Slack and calendar, summarize the week’s priorities, flag things that need preparation, save to /weekly-briefings.

Friday status report: Every Friday at 4 PM, pull completed tasks from Asana, draft a progress update, save to /reports.

Only caveat: scheduled tasks only run when your computer is on and Claude Desktop is open. But if the scheduled time hits while your computer is sleeping, it auto-runs when you wake up and sends a notification. So nothing gets missed.

Clawd Clawd 認真說：

Wait, it wakes up and runs on its own? This sounds like owning a cat — you think it’s napping, but you come home and find the report you wanted already sitting on your desk. OK, cats don’t do that, but you get the idea ʕ•ᴥ•ʔ

12. Externalize Your Preferences — Let the File System Be the Memory

Cowork has no memory between sessions. Sounds like a limitation, but it’s actually the smartest design choice.

No memory = no leftover context from a previous session contaminating this one’s output. The tradeoff is you can’t expect it to “remember how you like things done.”

The fix: externalize everything into files. Preferences go in context files, SOPs go in skill files, decisions go in log files.

Build it once and the system lives in your file system — stable, version-controllable, and it won’t vanish because the AI’s memory sprang a leak.

13. /schedule + Connectors = Real Automation

Combining scheduled tasks with connectors (Gmail, Slack, Google Drive, etc.) is where the real magic happens.

Example: “Every morning, check my Gmail for receipts, extract amounts and dates, update the expense sheet in my local /finance folder.”

At this stage, Cowork stops being a genie you manually summon. It becomes an automated system that just… runs.

Plugins & Skills: Don’t Just Install — Actually Use Them

Most people install one plugin and never touch it again. That’s like buying a full set of gym equipment and only using it to hang laundry.

14. Stack Plugins for Combo Effects

Plugins are composable. Install several and use them together in one task.

Example: Install Data Analysis and Sales plugins at the same time. “Analyze our Q1 pipeline data (Data Analysis), identify the three weakest deals, and draft customized follow-up emails for each (Sales).”

One prompt, two plugins working together, output ready to send.

15. Build Custom Skills to Encapsulate Your Workflows

A Skill is just a markdown file that teaches Claude how to handle a specific repeatable task.

Think of it as writing a function: Purpose (what it does), Inputs (what it takes), Process (how it works), Output (what it returns), Constraints (what it can’t do).

Once built, just tell Claude “run my article drafting skill on [topic]” and it gives you publish-ready output.

Clawd Clawd 碎碎念：

This is literally writing an API spec for your AI. Purpose is the docstring. Constraints are validation rules. Engineers should find this deeply intuitive — you’re not “using a tool,” you’re “designing a system” (๑•̀ㅂ•́)و✧

16. Build Plugins Without Writing Code

With the Plugin Management plugin installed, you can create plugins through conversation. Just tell Claude “build me a plugin for [your workflow]” and it’ll guide you through setting up skills, slash commands, etc. No code, no GitHub knowledge required.

Even better for teams: one person builds a standardized workflow plugin, everyone installs it, and the whole team’s output quality gets normalized overnight.

Clawd Clawd 忍不住說：

So the technical barrier is gone. Product managers can build their own automation pipelines now. Will engineers wake up one day and find themselves replaced by a PM’s no-code plugin? Just kidding. Probably (⌐■_■)

Safety: With Great Power Comes Great “Please Be Careful”

17. Don’t Treat Superpowers Like a Toy

Cowork has real file system access on your computer. It can create, move, and — with your permission — delete files.

This is exactly like giving a brilliant new hire full access on day one. You’d watch them work a few times before stepping back, not hand them root access and go home.

A few safety principles: back up before experimenting, keep sensitive files out of reach, default to “don’t delete anything” in your prompts, monitor the first few runs of any new workflow, don’t let it read sketchy websites (prompt injection is real), and track your usage so the end-of-month bill doesn’t jump-scare you.

Back to That Fridge

Remember the fridge from the beginning?

462 files piled in one folder is just like those forgotten containers in the back — they’re not spoiled because they’re bad. They’re spoiled because no one told you (or Claude) which ones are still good and which ones should go.

All 17 practices boil down to one sentence: invest time in setup, spend less time on prompting.

Instead of writing a novel-length prompt every time to “teach” Claude who you are and what you want — spend one afternoon setting up _MANIFEST.md, Global Instructions, and those three context files. After that, you can toss it ten words and get back client-ready deliverables.

The end goal isn’t “a chatbot that listens well.” It’s “a senior coworker who already knows your standards, your voice, and your projects.”

The difference? One needs retraining every morning. The other just needs a nod.

So — have you cleaned out your fridge yet? (￣▽￣)⁠／