02 · Knowledge worker with AI · Computer-Use

Daily tasks.

AI is your sparring partner for everyday work at the desk. Cowork is just one of several tools — here you learn what that looks like in practice.

Jump to the setup Start with the method

01 · Setup for knowledge workers

What you need to get started.

Eight items, done in around 30 minutes. First the Pro plan, the desktop app and the mobile link, then the briefing folder as context, then voice input as a bonus. Find more tools under [Tools].

Step 01
Claude Pro plan
The entry point. Pro at 20 USD/month opens Cowork, Projects, folder instructions and Computer-Use (which we cover later as a feature). Max 5x or 20x if you become a heavy user.
Pick a plan at Anthropic ↗
Step 02
Download the Claude desktop app
The native Mac/Windows app. A third window next to your browser and editor — Cowork lives inside it as its own tab. Drag and drop for files, faster switching between projects than the web tab.
Download Claude desktop ↗
Step 03
Create your markdown files
Three markdown files are the base for your briefing folder: about-me.md (your digital twin), schreibstil.md (how you sound), nicht-tun.md (what the AI must never do). One hour of effort, useful for years.
How to build the three files →
Step 04
Set up the briefing folder
Create a folder locally (e.g. ~/Documents/Briefing-Folder) and drop the three markdown files in. You give that folder to Cowork as folder instructions — and the same flow works in ChatGPT Projects or Perplexity Spaces.
Step 05
Save a default prompt for Cowork
In Cowork, set a default prompt once: „Read the markdowns from the briefing folder on every start and work from them.“ That way your context kicks in automatically without you calling it up every time.
Cowork deep dive →
Step 06
Wispr Flow for voice input
Dictation into any text field. Mac/Windows/iOS/Android. Roughly four times faster than typing — after a week you won't want to go back.
Get Wispr Flow ↗
Step 07
Optional, depending on need
More tools for specific use cases: ChatGPT Projects (persistent memory + folder), Perplexity Pro (cited-source research), Granola (local meeting notes), Raycast AI (hotkey AI on macOS), Gamma (deck generation).
Full tool cards under Tools →

Mantra

„Instead of writing the perfect prompt — let the AI ask you the right questions.“

That's the one sentence that changed everything for me. The AI doesn't know what's in your head. Stop asking it. Let it ask you.

01 · Sparring instead of prompting

The sparring method

A while back this would have been a list of prompting rules. Today one insight is enough: use the LLM as a sparring partner. Flip the direction. You describe what you want to achieve, the AI asks you all the questions it needs until it really understands your topic. While answering, you think the matter through yourself, instead of hand-crafting a perfect one-liner prompt.

Default sparring prompt (adapt and copy):

text·Sparring briefing

I want to work through this [task / topic].

Success criterion is: [xyz]

Ask me the right questions to gather all the information you need to complete the task successfully.

Always ask only one question at a time and wait for my answer.

If you still go the classic prompting route, three things make any prompt usable straight away.

1Step 1
Describe the goal
What should come out at the end? Which success criterion do you have in mind? If you can't say it in one sentence, the task is still too fuzzy.
2Step 2
Describe your approach
How would you tackle this as a human? Tell the AI. It takes it on step by step.
3Step 3
Define the output format
What should the result look like? Table, list, prose, slide? Giving examples is the strongest lever there is.

Sparring is the method. For it to really land, the AI needs one more thing: enough context about you.

02 · Briefing folder

Your briefing folder: three markdown files

Once you've created these three files, no AI starts from scratch with you again. You upload them into a Claude or ChatGPT project, or attach them to any chat. From then on the AI works with your context, your style and your rules, without you having to explain it all over again.

Markdown? Markdown is a simple text format with headings, lists and links. AI models read it perfectly. Exactly the right format for your briefing folder.

about-me.md

Your digital twin

What · A markdown file that captures you as a person: background, values, experience, strengths, thought patterns, decision logic.

Why · An AI with this file answers tasks from your perspective instead of a generic average mindset. It knows who's speaking and adapts tone, examples and recommendations.

How to build it

text·about-me.md

Please run a detailed interview with me to create a digital twin of me.

You should fully understand:
- my experience
- my values
- my perspective
- my communication style
- my expertise
- my way of thinking
- my decision patterns

Always ask only one question at a time and wait for my answer.

At the end, create a comprehensive markdown file that lets an AI take over all my tasks from my perspective.

Plan an afternoon for it. You'll use the result for years afterwards.

schreibstil.md

How you sound

What · A markdown file with your tone, sentence structure, favourite words, stylistic quirks. Plus a few original texts as reference.

Why · So texts from the AI sound like you, not like AI. Filler, sterile language and stock phrases disappear once the AI knows how your sentences are built.

How to build it

text·schreibstil.md

Please interview me and ask me questions that help you analyze my writing style. That includes me sharing texts I have written.

Then create a markdown file that mirrors my own writing style, including:
- tone
- sentence structure
- preferred words
- stylistic quirks

The file should also include my values and views on life, so they come through in my texts.

Tip: when you use this file, Claude Cowork currently matches your style most accurately.

nicht-tun.md

The do-not list

What · A markdown file with clear rules for what the AI must not do. Filler, disclaimers, sterile language, hallucinations, arbitrary bullet lists. Everything that annoys you lands here.

Why · If the AI doesn't know what you want to exclude, it does whatever. A do-not list prevents exactly that.

Examples from a typical file

No em dashes. The AI sprinkles them everywhere, humans barely use them. Use a period or a comma instead.
No AI filler like „let's dive in together“, „let's dive in“, „great that you ask!“, „a journey full of …“, „absolutely!“, „great question!“, „in today's fast-paced world“.
No empty superlatives: „revolutionary“, „groundbreaking“, „mighty“, „unmatched“.
When unsure: ask, don't hallucinate. A follow-up question beats an invented answer.

How to build it

text·nicht-tun.md

Help me create a do-not list for AI assistants — a markdown file with clear rules on what you should NOT do.

Ask me which typical AI patterns annoy me, for example:
- empty phrases
- disclaimers
- arbitrary lists
- meta commentary
- sterile language
- hallucinations

Ask only one question at a time, collect my examples, and at the end distill a file "do-not.md" that I can drop into every project.

What · Mission, product, audience, positioning, tone, taboo topics. One file that makes your brand tangible.

Why · As soon as you write emails, posts, pitches or texts for your company, the AI has the full brand logic in mind. No shaky descriptions, no off-brand voice.

How to build it

text·unternehmen.md

Interview me about my company and create a markdown file with:
- mission and vision
- product and positioning
- target audience
- tone and brand voice
- taboo topics and no-gos

Always ask only one question at a time.

These three files (plus optionally the fourth) are the briefing folder. From now on the AI delivers output that sounds like you.

Briefing folder · example

How the folder looks for a fictional person

Three markdown files, one shared folder. You upload it into Claude Projects, ChatGPT Projects or Gemini Gems and work from there with your own context.

about-me.md

Who you are

# About Me — Lena Berger

## Role
Product marketing manager at a B2B SaaS company in Munich.

## What I am good at
Explaining complex products simply. Running customer
interviews. Sharpening positioning.

## How I decide
Data first, then gut feeling. When neither is clear,
I talk to three customers.

## What matters to me
Clarity over coolness. Substance over hype.

schreibstil.md

How you sound

# My Writing Style

## Tone
Direct, warm, with self-irony. Never preachy.

## Sentence structure
Short, then long, then short. Fragments allowed.

## Favorite words
"concretely", "honestly", "at its core".

## Taboo
- no empty phrases
- no marketing adjectives
- never "in today's times"

nicht-tun.md

What the AI avoids

# Do-Not List

- No em-dashes.
- No disclaimers like "it depends, of course".
- No hype adjectives (revolutionary, groundbreaking).
- No arbitrary lists of three when only two points fit.
- When unsure: ask, do not hallucinate.
- Never make up numbers.

Note · These three example files are made up but realistic. Your own ones come from the default prompts in the cards above.

03 · Modes & features

Modes and features worth knowing

In 2026, LLMs are no longer chat windows. They research autonomously, see images, click inside your apps, listen to you and remember projects for weeks. Here are the seven modes a knowledge worker uses every day.

Mode 01

Research (Deep Research)

What · Autonomous, multi-step web research. The AI scans dozens of sources, summarises and cites.

What for · Market analyses, competitor comparisons, background briefings before calls.

Start deep research · 12 sources checked

Available in: ChatGPT, Claude, Gemini, Perplexity, Grok

Mode 02

Vision (image understanding)

What · You upload an image, screenshot, PDF or whiteboard photo. The AI reads, describes and analyses.

What for · Transcribe whiteboard photos, read out charts, UI reviews, extract receipts.

Drag and drop · image thumbnail

Available in: ChatGPT, Claude, Gemini

Mode 03

Computer-Use

What · The AI sees your screen, clicks, types and scrolls inside your apps. Like a virtual co-worker.

What for · Email triage, tidying folders, recurring workflows in local apps.

Desktop capture · AI working

Available in: Claude (Computer-Use + Dispatch), ChatGPT (Agent), Gemini

Best to set up on a separate machine without sensitive data.

Mode 04

Voice

What · You speak, the AI listens. Either as natural conversation or as dictation into any text field (Wispr Flow).

What for · Dictating emails, brainstorming on a walk, prompt input without a keyboard.

Mic pulse · waveform

Available in: ChatGPT Advanced Voice, Gemini Live, Grok, Wispr Flow for any text field

Mode 05

Projects

What · Persistent workspace with custom instructions, uploaded files, shared chat history. At Claude it's called Projects, at OpenAI also Projects, at Google it's Gems.

What for · Recurring workflows (quarterly reports, contract review, competitor analysis) where the context stays the same.

Folder · about-me.md · schreibstil.md · nicht-tun.md

Available in: Claude Projects, ChatGPT Projects (up to 40 files), Gemini Gems

Mode 06

Agent mode

What · The AI runs multi-step tasks on its own. Browser, mail, calendar, spreadsheets. You give the direction, it does the work.

What for · „Plan a trip and book the flights“, „analyse three competitors and build a deck“, morning email triage.

Task list · checkmarks turn green

Available in: ChatGPT Agent (40-400 tasks/month depending on plan), Claude Dispatch, Gemini

Mode 07

Connect your phone (optional)

What · Pair Cowork or ChatGPT with your phone and bring the AI on the go. With Claude it's Dispatch — you hand the model a task on the road and it runs on your desktop.

What for · Quick sparring runs on the way to a meeting, dictating briefings, kicking off tasks without opening the laptop.

iOS/Android app · microphone · sync to desktop

Available in: Claude Mobile + Dispatch, ChatGPT Mobile, Gemini Mobile

Which modes you actually need depends on the stack you work with. Exactly what the next section is about.

05 · Cowork deep dive

Why Cowork is our default — and when to pick something else.

Cowork is Anthropic's knowledge-worker app. We recommend it as the default, though it isn't the only way.

What Cowork can do

Third tab

Its own tab next to chat and code in the desktop app. Persistent sessions per project.

Folder instructions

about-me.md, schreibstil.md, nicht-tun.md load automatically — your briefing folder.

Scheduled tasks

Run tasks at fixed times — weekly reports, daily briefings, monitoring.

Dispatch (mobile)

Trigger Computer-Use from your phone. Set up the task, the desktop runs it.

Industry bundles

Legal, small business, marketing ops — preconfigured skills and prompts per role.

When to pick something else

Cowork is the default recommendation. Two alternatives that fit better in specific setups:

ChatGPT Agent

When: When your team already lives in the OpenAI ecosystem and browser automation is the main focus.

OpenAI's Computer-Use product in the Plus/Pro plan. Strong on clicking through web apps, weaker on local desktop and markdown discipline.

Provider ↗

Perplexity Comet

When: When the focus is on research.

Comet browser with built-in AI, multi-model. Strong on sources, weaker on long-form writing.

Provider ↗

04 · As of May 2026

The Computer-Use tools compared

Four tools that really do Computer-Use — steering the desktop or browser autonomously. We curated the full knowledge-worker stack (Wispr Flow, Granola, Notion AI, Raycast, Mem …) on the Tools page.

Job-to-be-done · quick pick

You want to …	Take	Why
Let an AI drive the desktop	Claude Cowork (+ Dispatch for the phone)	Clicks inside local apps, controllable from mobile.
Let an AI drive the browser autonomously	ChatGPT Agent	OpenAI's Computer-Use product on the Plus and Pro plan.
Do web research in an AI browser	Perplexity Comet	AI browser with cited sources, multi-model pick.
Build your own browser agent in code	Browser-Use	Open-source Python framework on Playwright.

The four Computer-Use tools in detail

Per card: tagline, what, pricing, strengths, weaknesses, link.

Claude Cowork (with Dispatch)

Claude's Computer-Use — as a third tab in the desktop app. Dispatch drives it remotely from your phone.

What · Cowork runs in the Claude desktop app, has access to the file system, plans tasks, drives local apps. Dispatch is the mobile companion: QR scan on your phone, done — you start at the desktop, check in and give follow-up commands on the go.

Pricing · Included in the Pro plan ($20/mo), Max 5x $100/mo, Max 20x $200/mo

Strengths

Computer-Use + workspace + mobile remote control in one product.
Briefing folder linked directly inside the workspace.

Weaknesses

Multi-app tasks currently succeed reliably around 50 % of the time.
Computer-Use is in research preview.

claude.com/product/cowork

ChatGPT Agent

OpenAI's Computer-Use product on the Plus/Pro plan.

What · A dedicated agent mode in ChatGPT that takes over the browser autonomously — researches, clicks, fills forms, plans tasks. Mechanics comparable to Claude Cowork, with a different feature scope and a different ecosystem.

Pricing · Plus $20/mo (limited agent quota), Pro $200/mo (larger quota), Business $25/seat

Strengths

Deep integration with the ChatGPT ecosystem (custom GPTs, connectors, Projects).
Good for structured research and booking tasks.

Weaknesses

Sandboxed browser — no access to local desktop apps.
Quotas per plan, no multi-user setup under Plus.

chatgpt.com

Perplexity Comet

Perplexity's AI browser with Computer-Use.

What · A dedicated browser from Perplexity. Cited-source research, deep research, Computer-Use actions directly in the tab. Multi-model pick (Claude, GPT, Gemini).

Pricing · Included with Perplexity Pro — $20/mo (or $200/year ≈ $16.67/mo), Max $200/mo

Strengths

AI browser with source tracking — research stays verifiable.
Multi-model selection per request.

Weaknesses

A separate browser is required — if you live in your main browser, you'll have two windows open.
Weaker on writing with style than Claude or ChatGPT.

www.perplexity.ai/comet

Browser-Use

Open-source framework for your own browser agents.

What · Python framework on Playwright (79k GitHub stars). The DIY alternative to ChatGPT Agent. You build your own workflows and only pay LLM API calls.

Pricing · Framework is free, you only pay LLM API calls. Browser-Use Desktop is also open source.

Strengths

89.1 % success rate on the WebVoyager benchmark (SOTA).
Full control, local execution, no vendor lock-in.

Weaknesses

Python comfort required.
No one-click setup for non-devs.

github.com/browser-use/browser-use

Full tool cards under Tools

05 · Best practices

Your first briefing folder in 5 steps

Theory is nice, the effect kicks in once you've actually built the folder. Plan an afternoon for it, ideally with coffee and no calendar appointments. Afterwards every future AI task runs at a different level.

Pick an LLM with a project feature
Claude (Projects or Cowork) or ChatGPT (Projects). Both work. If you already have an account, stick with it.
Create a new project called „My briefing“
Leave it empty for now, you'll fill it in a moment.
Run the three interviews one after another
First about-me.md, then schreibstil.md, then nicht-tun.md. Each with the prompts from section 2. About 30 to 60 minutes per interview. You answer questions, the AI builds the file.
Upload the three finished markdown files into your project
In Claude under „Project Knowledge“, in ChatGPT under „Files“. Optionally also unternehmen.md if you have one.
Test task with the sparring prompt
Take the default prompt from section 1, describe a real task (an email, a LinkedIn post, a research question), and let the AI ask you. You'll feel the difference right away.

What happens once the folder is in place

From now on every output sounds like you. The AI no longer has to guess who is asking and in which tone to answer. You save the first 5 to 10 minutes of context-setting per chat, every single day, starting today.

Recommendations for this topic

Who we follow specifically here.

Creators and newsletters that fit this cluster thematically. Outside voices, curated by us.

Ruben Hassid

One of the clearest LinkedIn voices on AI for everyday knowledge work. Concrete workflows, honest reviews, no hype.

Language:: EN

Take a look Instagram Substack

Ruben Hassid · Substack

His Substack newsletter with deep dives on AI in daily life — where it actually lands, where it still struggles, what's coming next.

Frequency:: weekly
Language:: EN

Take a look

All recommendations

On to the next station

Build products with AI

From concept to finished app. With Claude Code, MCPs, skills and token economy.

Take a look

Daily tasks.

Claude Pro plan

Download the Claude desktop app

Create your markdown files

Set up the briefing folder

Save a default prompt for Cowork

Wispr Flow for voice input

Optional, depending on need

Describe the goal

Describe your approach

Define the output format

Your digital twin

How you sound

The do-not list

Research (Deep Research)

Vision (image understanding)

Computer-Use

Voice

Projects

Agent mode

Connect your phone (optional)

What Cowork can do

When to pick something else

Job-to-be-done · quick pick

The four Computer-Use tools in detail

Claude Cowork (with Dispatch)

ChatGPT Agent

Perplexity Comet

Browser-Use

Pick an LLM with a project feature

Create a new project called „My briefing“

Run the three interviews one after another

Upload the three finished markdown files into your project

Test task with the sparring prompt

Who we follow specifically here.

Ruben Hassid

Ruben Hassid · Substack