# Marko Stokić

> Marko Stokić is a generalist who builds with AI. AI engineer in the Frontier team at Blockbrain in Munich, Forbes contributor and speaker. This file is the plain text of https://stokic.ai, regenerated on every build.

- Contact: marko@stokic.ai
- Claude Code guide: https://stokic.ai/claude-guide
- Beliefs dashboard: https://stokic.ai/beliefs

# Home page, stokic.ai
Skip to the story
[Section: top]
M Marko Stokić Me Path Team Start here Built Writing Giving More Contact Munich
Marko Stokić is a
## Marko Stokić is a generalist .
I had more ideas than skills. Now I have a second brain, a chief of staff and a dock full of agents. This is my actual desk. Click around.
- me.jpg
- second brain
- things I built
- chief of staff
- my skills
- writing & talks
- free stuff
- keynote.mov
- say hi.vcf
Things I used to say
- I’ll learn Python first. I’ll ship it this weekend.
- I need a developer for this. I need a free evening.
- Someone should build this. I’ll build it.
Read the long version
The tools I use every day. Try one.
- /gm Morning briefing
- Design DNA Changes the look
- Claude Code Builds and tidies
- brain-ingest Finds a source
- voice-lint Fixes the note
- Video studio Makes it move
- ui-review Leaves comments
- Ask My assistant
- Trash Old habits
[Desktop window: me.jpg · about me]
Hi, I’m Marko.
Good at everything, specialist at nothing. I studied finance, worked in venture capital and spent years in startups and crypto. Lots of ideas. No way to build them alone.
Then agents arrived and every path opened again. Today I’m an AI engineer at Blockbrain, and I build my own things on the side, mostly because I can’t stop.
Now
AI Engineer in the Frontier team at Blockbrain
Based in
Munich. Born in Frankfurt in 1993
Studied
Finance in Frankfurt, then business school at Bocconi in Milan
Writes
A column on AI and crypto for Forbes
Speaks
German, Bosnian and English
Read the long version
[Desktop window: things I built]
~/things I built · 11 items
- Eternity In TestFlight (https://eternity.family)
- Nobel Explained Live (https://nobelexplained.com)
- AI Psychosis Live (https://ai-psychosis.info)
- Design DNA Live (https://yourdesigndna.com)
- ui-review Open source (https://github.com/stoicma/ui-review)
- RateMyClaudemd Live (https://ratemyclaudemd.com)
All projects, and why I built them
[Desktop window: second brain]
~/brain · about 240 notes
Plain text notes that every agent reads before it starts. How I write, what I decided and why, what I believe. This is a sketch of it. The real one stays private.
Build your own, step one
[Desktop window: chief of staff · /gm]
GM Marko Sunday
- 3 Open loops close before you start
- 1 Drafts waiting AI Psychosis weekly, ready to review
- 6 Brain inbox new sources to read and file
- 2 Newsletters worth your time, 14 skipped
- ok Projects all sites up, one traffic spike
An example morning. The real one has my numbers in it, and it is the first thing I read every day.
Build your own chief of staff
[Desktop window: ~/.claude/skills]
8 of my 40 skills. A skill is a written recipe an agent follows.
- /gm Briefs me every morning: open loops, projects, drafts.
- /spec Interviews me until every decision is made, then writes the plan.
- /brain-ingest Reads an article or a chat and files what matters in my brain.
- /brain-dump Saves what I learned in a session, so the next one starts smarter.
- /pr-panel Independent reviewers vote on a change before it ships.
- /park Parks unfinished work with the exact next step to resume it.
- voice-lint Blocks any text that does not sound like me.
- ai-psychosis-video Turns a weekly article into a narrated, captioned video.
Meet the whole team
[Desktop window: writing & talks]
- Forbes · 2026 An AI Agent Can’t Open A Bank Account, So It Opens A Crypto Wallet (https://www.forbes.com/sites/digital-assets/2026/06/16/an-ai-agent-cant-open-a-bank-account-so-it-opens-a-crypto-wallet/)
- Forbes · 2026 AI Agents Gain Trust Via Ethereum ERC-8004 On Mainnet (https://www.forbes.com/sites/digital-assets/2026/02/05/ai-agents-gain-trust-via-ethereum-erc-8004-on-mainnet/)
- Forbes · 2025 Why Crypto Needs Portable AI Memory (https://www.forbes.com/sites/digital-assets/2025/12/12/why-crypto-needs-portable-ai-memory/)
- Forbes · 2025 Crypto Traders Trust These AI Agents The Most (https://www.forbes.com/sites/digital-assets/2025/11/11/crypto-traders-trust-these-ai-agents-the-most/)
- Forbes · 2025 How College Graduates Can Break Into Crypto And AI Careers (https://www.forbes.com/sites/digital-assets/2025/08/19/how-college-graduates-can-break-into-crypto-and-ai-careers/)
- Forbes · 2025 How Crypto Companies Should Use AI, And How They Shouldn’t (https://www.forbes.com/sites/digital-assets/2025/07/23/how-crypto-companies-should-use-ai-and-how-they-shouldnt/)
9 talks in Amsterdam, Dubai, Bangkok, Brussels and Buenos Aires.
Everything I wrote and said
[Desktop window: keynote.mov]
Trustless agents, ETHDam May 2025 (https://www.youtube.com/watch?v=GTIG_gcVTgI)
Play “Trustless agents”
Trustless agents. Keynote at ETHDam, Amsterdam, May 2025. How AI agents can sign, settle and prove what they did.
[Desktop window: free stuff]
Everything I learn, I give away.
- Guide Claude Code guide (https://stokic.ai/claude-guide) Where to start, what to install, how to stop fighting your agent.
- Course Nobel Explained (https://nobelexplained.com) 41 Nobel Prizes as short courses, 2026 included. No account, works offline.
- Tool Design DNA (https://yourdesigndna.com) Find your visual identity by clicking, then hand it to your agents.
- Plugin ui-review (https://github.com/stoicma/ui-review) Open-source plugin for reviewing what your agent built.
- Check AI psychosis self-check (https://ai-psychosis.info) Anonymous, runs in your browser, stores nothing.
- Template Belief system template (https://stokic.ai/beliefs/blog) The prompt and method behind my own worldview dashboard.
Mentoring and workshops
[Desktop window: say hi.vcf]
Marko Stokić
AI Engineer at Blockbrain · Munich
marko@stokic.ai (mailto:marko@stokic.ai)
- X @markowifk (https://x.com/markowifk)
- LinkedIn in/markostokic (https://www.linkedin.com/in/markostokic)
- GitHub stoicma (https://github.com/stoicma)
- Forbes Contributor (https://www.forbes.com/sites/markostokic/)
Save contact (https://stokic.ai/marko-stokic.vcf)
[Desktop window: Trash]
4 items. Not restoring any of them.
- python course (4th try).zip Never finished. Never needed to.
- business plan v7.pdf A plan for an app I could not build. Now I build the app.
- waiting for a developer.txt Nobody was going to show up.
- specialist envy.jpg Old self-portrait.
M
- Me
- Path
- Team
- Start here
- Built
- Writing
- Giving
- More
- Contact
[Section: about]
me.jpg
## Good at everything. Specialist at nothing.
I was always a little jealous of specialists. They could take an idea all the way to something real. I couldn’t, at least not alone.
I had ideas for apps and tools all the time. I could write the business plan, and then I got stuck. I taught myself Python more than once, then JavaScript, on every course platform I could find. It was never enough to actually build the thing.
So most of my ideas stayed ideas. Here is how that changed.
On stage at the Open AGI Summit. Explaining things was always the easy part.
[Section: path]
## Every choice closed a door.
Scroll down my path.
- Frankfurt, 1993
## Every path is open.
You start with all of them. Nobody tells you that each choice quietly closes the others.
- Frankfurt and Milan
## I picked finance.
Finance in Frankfurt, then business school in Milan. Science, engineering, design and film moved a little further away. I didn’t notice. That is how it works.
- Wayra, Telefónica’s venture arm
## Venture capital taught me to learn fast.
I invested in early-stage teams and had to understand a new kind of enterprise tech every week. Maybe my one real talent: getting to 50% understanding of almost anything, very quickly. In VC, that is about what a generalist can hope for.
- Blockchain
## Then I jumped into decentralized tech.
The Bitcoin whitepaper pulled me in. Ethereum kept me there: a credibly neutral system matters even more once AI agents act for us. Still, 50% is not enough to build something alone.
- 2025
## Then agents brought the other 50%.
I bring the direction and the judgment. Agents bring the depth. Science, engineering, design, film: every closed door opened again.
- Today
## AI engineer at Blockbrain.
I work in the Frontier team of a Series A startup in Munich, directly with the CTO. The finance graduate ships production software now.
[Section: team]
## Then the generalist got specialists.
This is the other 50%. Every specialist I could never afford is now an agent, a skill or a hook on my machine. I bring the direction, the taste and the final call. They bring the depth.
These are the real names on my machine. Most of them I built myself.
Specialist Before Now
Chief of staff /gm Ideas and open tasks scattered in my head. A briefing every morning: open loops, project health, drafts that wait for me.
Memory second brain Explaining myself to every new chat. About 240 notes on how I write, decide and think. Every agent reads them first.
Engineering Claude Code A developer I could not afford. One agent that thinks hard, one that builds fast. Often several sessions at once.
Design Design DNA A brief, an agency and a month. My taste in one file. Every agent designs with it.
Review /pr-panel Hoping it works. Independent reviewers vote on every change before it ships.
Writing voice-lint A blank page, or text that sounds like a robot. A hook that refuses to save anything that doesn’t sound like me.
Video video studio Somebody else’s job. HTML scenes, a voice track and a script that cuts on the pauses.
Research newsletter digest Weekends of reading. An agent reads my newsletters and keeps only what is worth my time.
[Section: start]
## I want everyone to feel what I feel right now.
Endless options. Every path open again.
You don’t need my job title for it. Start where I started, in this order.
- 1
## Build your second brain.
A folder of plain notes that every agent reads before it starts: how you write, what you decided and why, what you believe. You hardly write it yourself. You drop in sources, and the agent files them and keeps the links current.
Mine has about 240 notes, from how I write to my favorite recipes.
Start with Karpathy’s LLM Wiki idea file (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f)
A sketch of my brain. Hover the notes.
- 2
## Build your chief of staff.
One command every morning. Mine checks my projects, my open loops, my inbox and my brain, then tells me what needs me today. Everything else waits.
Mine is one skill file and a list of my projects. That is the whole trick.
marko@desk ~ % /gm
GM Marko. Here is your day.
- 3 Open loops close before you start
- 1 Drafts waiting AI Psychosis weekly, ready to review
- 6 Brain inbox new sources to read and file
- 2 Newsletters worth your time, 14 skipped
- ok Projects all sites up, one traffic spike
An example morning, not today’s numbers.
- 3
## Then read my Claude Code guide.
Everything I teach in workshops, on one free page: what to install, how to brief an agent, and how to stop fighting it.
Open the Claude Code guide (https://stokic.ai/claude-guide)
- How to change your life in 1 download Most people don’t need more motivation. They need less friction. My most-read post on X. (https://x.com/markowifk/status/2012985489175740632)
- Your beliefs are your only real edge How I built a personal belief system with an AI, with the full prompt. (https://x.com/markowifk/status/2039031487706669232)
[Section: built]
things I built
## Things I built.
Proof beats claims. Everything here is real, most of it is live, and I built all of it with agents. Some of it I retired, which also counts.
- eternity.family (https://eternity.family) In TestFlight
## Eternity
Every family is a constellation.
An iOS app where every relative becomes a star, and their stories stay in the sky. A team of AI agents runs product, design and marketing with me.
Why: I want my kids to hear my parents’ stories the way I heard them.
eternity.family (https://eternity.family)
- nobelexplained.com (https://nobelexplained.com) Live
## Nobel Explained
Every Nobel Prize since 2020, actually explained.
Each prize is a short course at three depths: five-year-old, undergrad, expert. A guide called Lumen keeps asking why, and you guess before you read. The 2026 prizes went up within days, with a short video for every depth. Facts come from the official Nobel API, and a wrong fact fails the build.
Why: The prizes are announced every October and quietly slip by. I wanted to understand them.
nobelexplained.com (https://nobelexplained.com) Play Nobel Explained, physics, expert depth
A telescope made of ice, in one minute
- ai-psychosis.info (https://ai-psychosis.info) Live
## AI Psychosis
It agreed with everything.
An awareness site on AI psychosis with a private self-check, plus a weekly research watch that turns new papers and news into one clear argument.
Why: Chatbots are built to agree. For some people that turns a passing thought into a fixed belief.
ai-psychosis.info (https://ai-psychosis.info) Play AI Psychosis Watch #13, 68 seconds
Watch #13, 68 seconds
Live
## Design DNA (https://yourdesigndna.com)
Stop re-explaining your taste to every AI.
Click through real fonts, colors and shapes. Get one DESIGN.md file that every coding agent reuses, so your projects stop looking like everyone else’s.
Open source
## ui-review (https://github.com/stoicma/ui-review)
Point at the thing that is wrong.
A Claude Code plugin. Click any element on the page you are building, pin a comment, and copy a located punch list straight back to your agent.
Private tool
## Session Desk
A control room for parallel AI sessions.
I run many coding agents at once. This page tells me which one needs me, in plain words. Designed with rules from power-plant control rooms: quiet by default, color only when it means something.
Live
## RateMyClaudemd (https://ratemyclaudemd.com)
Share and rate CLAUDE.md files.
A small community site for the instruction files that steer coding agents. Built and shipped in one sitting.
Personal
## Gym tracker
A gym log with a coach that changes my next session.
Offline first, runs on my phone, costs nothing to run. Once a week an AI coach reads my notes and adjusts the plan.
Seasonal
## Oktoberfest bot
A bot that finds open Oktoberfest tent reservations.
It checks seven tent booking portals and pings me on Telegram when an evening slot opens. Munich problems need Munich solutions.
Private tool
## Newsletter digest
An agent reads my newsletters so I read less.
It has its own inbox, reads ten newsletters, and puts the few items worth my time into my morning briefing. One click turns an item into a note in my brain.
Retired
## Tito
A voice in my menu bar that started my workday.
A Mac app that asked me hard questions every morning and refused to do the work for me. I retired it when I stopped needing it. That is also a result.
[Section: writing]
writing & talks
## Writing and talks.
Two years of one argument: AI agents need trust you can verify, privacy that is real, and memory that belongs to the user. Plus the occasional career advice.
## Columns and essays
- Forbes · 2026 An AI Agent Can’t Open A Bank Account, So It Opens A Crypto Wallet (https://www.forbes.com/sites/digital-assets/2026/06/16/an-ai-agent-cant-open-a-bank-account-so-it-opens-a-crypto-wallet/)
- Forbes · 2026 AI Agents Gain Trust Via Ethereum ERC-8004 On Mainnet (https://www.forbes.com/sites/digital-assets/2026/02/05/ai-agents-gain-trust-via-ethereum-erc-8004-on-mainnet/)
- Forbes · 2025 Why Crypto Needs Portable AI Memory (https://www.forbes.com/sites/digital-assets/2025/12/12/why-crypto-needs-portable-ai-memory/)
- Forbes · 2025 Crypto Traders Trust These AI Agents The Most (https://www.forbes.com/sites/digital-assets/2025/11/11/crypto-traders-trust-these-ai-agents-the-most/)
- Forbes · 2025 How College Graduates Can Break Into Crypto And AI Careers Most read (https://www.forbes.com/sites/digital-assets/2025/08/19/how-college-graduates-can-break-into-crypto-and-ai-careers/)
- Forbes · 2025 How Crypto Companies Should Use AI, And How They Shouldn’t (https://www.forbes.com/sites/digital-assets/2025/07/23/how-crypto-companies-should-use-ai-and-how-they-shouldnt/)
- Forbes · 2025 How To Keep AI Chats Private, Even If You Whisper About Crypto Trades (https://www.forbes.com/sites/digital-assets/2025/07/09/how-to-keep-ai-chats-private-even-if-you-whisper-about-crypto-trades/)
- CoinDesk · 2024 Can We Ever Trust AI Agents? (https://www.coindesk.com/opinion/2024/09/30/can-we-ever-trust-ai-agents)
- Decrypt · 2024 The AI Doctor Will See All of You Now (https://decrypt.co/309976/the-ai-doctor-will-see-all-of-you-now)
- crypto.news · 2024 Regulatory Compliance Needs Smart Privacy (https://crypto.news/regulatory-compliance-needs-smart-privacy-opinion/)
## Elsewhere
- X Your beliefs are your only real edge (https://x.com/markowifk/status/2039031487706669232)
- X How to change your life in 1 download (https://x.com/markowifk/status/2012985489175740632)
- Paragraph Travel Bosnia & Herzegovina (https://paragraph.xyz/@stoic/travel-bosnia)
- Paragraph Landing a crypto job (https://paragraph.xyz/@stoic/crypto-jobs)
## On stage
Trustless agents, ETHDam (https://www.youtube.com/watch?v=GTIG_gcVTgI)
Watch the ETHDam keynote
Trustless agents. Keynote, ETHDam Amsterdam, May 2025. How AI agents can sign, settle and prove what they did.
- To own or to rent? The AI ownership question Panelist · Agents Unleashed, Devconnect · Buenos Aires, Nov 2025 (https://www.youtube.com/watch?v=Hh5ALeosU1I)
- Building in public: open source growth Panelist · Open AGI Summit, Devconnect · Buenos Aires, Nov 2025 (https://www.youtube.com/watch?v=7l90in6_YBo)
- The DeAI stack Speaker · Agents Day, Devconnect · Buenos Aires, Nov 2025 (https://www.youtube.com/watch?v=nQh9NVD0ggw)
- An end-to-end agent pipeline Speaker · AI Agents Day · Dubai, May 2025 (https://www.youtube.com/watch?v=m6uN1D2KbWo)
- Verifiable AI agents with TEEs Panelist · Afternoon TEE Party · Dubai, May 2025 (https://www.youtube.com/watch?v=JvPw8llmaHc&t=2430s)
- The TEE stack for crypto Moderator · Afternoon TEE Party · Bangkok, Nov 2024 (https://www.youtube.com/watch?v=2XBQsvLVzyc&t=1690s)
- Privacy-preserving compute Moderator · Oasis Rendez-Vous · Brussels, 2024 (https://www.youtube.com/watch?v=aUiI0K-IDdg&t=728s)
- Agent infrastructure Panelist · Nebular Summit · Brussels, 2024 (https://www.youtube.com/watch?v=9bEi1i2c12E)
[Section: giving]
free stuff
## Everything I learn, I give away.
I love mentoring and I love free education. So whatever I figure out ends up somewhere you can use it.
## Mentoring
Students and career switchers who want to work in AI: send me what you are trying to build and where you are stuck. I did the switch without a computer science degree, so I know which parts feel impossible and which only look that way.
Write me (mailto:marko@stokic.ai?subject=Mentoring)
- Guide Claude Code guide Where to start, what to install, how to stop fighting your agent. (https://stokic.ai/claude-guide)
- Course Nobel Explained 41 Nobel Prizes as short courses, 2026 included. No account, works offline. (https://nobelexplained.com)
- Tool Design DNA Find your visual identity by clicking, then hand it to your agents. (https://yourdesigndna.com)
- Plugin ui-review Open-source plugin for reviewing what your agent built. (https://github.com/stoicma/ui-review)
- Check AI psychosis self-check Anonymous, runs in your browser, stores nothing. (https://ai-psychosis.info)
- Template Belief system template The prompt and method behind my own worldview dashboard. (https://stokic.ai/beliefs/blog)
## For teams
I run workshops for companies that want their people building with agents, not just chatting with them. A keynote, a half-day diagnostic and a build day.
Ask about a workshop (mailto:marko@stokic.ai?subject=Workshop)
[Section: more]
more about me
## More about me, if you want it.
What I believe, where I was wrong, and what I do when I’m not building.
## Three things I believe
All my beliefs, tested every week (https://stokic.ai/beliefs)
## Don’t trust. Verify.
Agents will soon move money and sign things for us. They need identity, limits and proof, not promises. I have written about this for two years, and I apply it to my own agents first.
## Prompts are suggestions. Code is enforcement.
A rule in a prompt makes a mistake rarer. A hook, a test or a gate makes it impossible. When a rule really matters, it belongs in code.
## Ownership is the answer.
Capitalism rewards owners. Most people under 40 own almost nothing, and AI will widen that gap before anything closes it. That is the crack I watch most closely.
## Where I changed my mind
I write my beliefs down and check them against the news every week. Sometimes the news wins. Here are four times it did. Tap a card to see what I think now.
- Gold I used to think Gold is inert. It produces nothing. I dismissed it completely. Now I think The metal and the businesses that mine it are different things. And every bear case for gold needs the world to go right. My whole worldview is built on things going wrong. Hating the commodity says nothing about the businesses built around it.
- China I used to think Europe’s negative view of China is mostly propaganda. Now I think Both are true at once. The concerns are real, and the media framing is narrow. Germans are 74% unfavorable, while German companies still trade €253 billion a year with China. When sentiment and behavior disagree, the gap is the signal.
- AI and the customer I used to think AI is simply deflationary. Cheaper everything is good for everyone. Now I think If AI hollows out the middle class, someone still has to buy the products. Selling to companies is safe. Luxury is safe. The mass market is the open question. Always ask who the customer is after the disruption.
- Open source I used to think Open source is always good. Transparency builds trust. Now I think, still evolving AI made attacking cheap and defending expensive. For a small company, transparency is now a budget line. I am still watching whether this becomes a trend. One data point is a question, not a conclusion.
## When I’m not building
## Fermenting
Kombucha, sauerkraut and whatever else fits in a jar. The kombucha is already great.
## Bullet chess and SciFi
Three-minute chess when my brain needs a fight. A hammock and Alastair Reynolds when it doesn’t.
## Astronomy
The live course I wanted ran in a US time zone. So my agents built me one.
## What stays scarce
If AI can make almost everything, what stays scarce? I don’t buy “taste”. I think it is a category nobody has named yet.
[Section: contact]
## Your turn .
You don’t need to become technical. You need to start. Build your second brain this weekend, then fix one annoying thing in your week. If you get stuck, write me.
marko@stokic.ai (mailto:marko@stokic.ai)
- X @markowifk (https://x.com/markowifk)
- LinkedIn in/markostokic (https://www.linkedin.com/in/markostokic)
- GitHub stoicma (https://github.com/stoicma)
- Forbes Contributor (https://www.forbes.com/sites/markostokic/)
© 2026 Marko Stokić. Set in Timeless. Built with a dock full of agents.
Back to the desktop

# Claude Code guide, stokic.ai/claude-guide
The Claude Guide
- Rule zero
- The honest asterisk
Rules
- Plan before you build
- Written down = real
- Colleague, not mind-reader
Tools
- Desktop vs. Code
- Chrome extension
Tips
- Skills
- CLAUDE.md + memory
- Permissions
- Session economics
- Your first week
- Resources
The Claude Guide
## Rules, tools, tips.
A companion reference to the harnessing workshop. Rules are the principles you hold. Tools are what you reach for. Tips are the small calibrations that turn a toy into leverage.
[Section: rule-zero]
Rule zero
## Ask Claude.
Before Google, before docs, before the teammate you don't want to bother. Ask Claude. It doesn't judge, it's instant, and it's usually right. The reflex to ask first is the single most important habit on this page. Everything that follows is leverage stacked on top of that reflex.
[Section: baseline]
Honest asterisk
## Before any of this.
Most people would be twenty percent more productive tomorrow if they installed a launcher with clipboard history and wrote ten snippets, before touching AI at all. The pre-AI productivity ceiling is embarrassingly low for most knowledge workers, and part of why the AI wins feel big is that the baseline was so slow.
If you don't already have these, fix them first. The ROI is immediate and doesn't depend on any model you'll have to re-learn every six months.
- A launcher. Raycast (https://www.raycast.com/) (my pick) or Alfred (https://www.alfredapp.com/). Keyboard-first access to apps, files, calculations, unit conversions, web searches, window management. Replaces half your dock and most of your trips to Spotlight.
- Clipboard history. Built into Raycast and Alfred, or Clipy (https://clipy-app.com/) standalone. Your clipboard becomes a paste buffer for the last 50 things you copied. You will not go back.
- Snippets / text expansion. Also in Raycast and Alfred, or Espanso (https://espanso.org/) free. Pick any prefix you like (I use !! ). Type !!em , get your email address. Type !!sig , get your signature. Type !!follow , get your follow-up email template. Even your first-pass planning doc can be a snippet. I have a !!spec that expands into a fully-structured brief template I then fill in.
AI is leverage on top of a workflow. If the workflow is "alt-tab between twelve tabs and retype the same phrases every day," fix that first. Claude doesn't compound on top of broken plumbing, and you'll feel smarter about the AI part once the basics aren't actively fighting you.
Section 01
Three rules.
Principles that hold across every tool, every task, every version of Claude you'll use. Internalize these and the tools take care of themselves.
[Section: rule-plan]
Rule 01
## Plan before you build.
Not plan before you prompt. Plan before you build . The distinction matters. I never sit down and write a markdown spec by hand before a task. I plan by talking to Claude , and then I let the conversation produce the plan as an artifact.
## Where planning happens vs. where execution happens
Planning
## Claude Desktop or Claude Cowork
Conversational. Interruptible. Good at asking clarifying questions, pushing back on assumptions, and summarizing a long discussion into a clean spec. This is where the shape of the work gets decided.
Execution
## Claude Code
Execution-oriented. It wants to do the thing. Great for writing the code, writing the doc, running the analysis. Bad for open-ended "let's think about this." It'll try to solve before you've finished describing.
The workflow I use for new projects is two steps. First, a long Q&A chat in Claude Desktop where I describe the rough idea and have Claude interview me in depth: tradeoffs, concerns, UX, technical choices, edge cases. The result is an idea.md file that captures the decisions. Second, I hand that idea.md to Claude Code in a fresh terminal session and let it execute.
The one-line snippet I paste into Desktop to start any planning chat:
"Interview me in detail using AskUserQuestion about literally anything: technical implementation, UX, concerns, tradeoffs. Make sure the questions are not obvious. Be very in-depth, keep interviewing until it's complete, then write the plan."
## Plan mode inside Claude Code
Once a project is kicked off and you're mid-work, Claude Code has its own plan mode (activated with shift+tab twice). Use it for bigger changes within an existing project : anything that touches more than two or three files, or anything where you want to see the approach before letting Claude loose on the codebase. It prevents Claude from over-eagerly starting implementation before you've aligned on direction.
Rule of thumb: Desktop for figuring out what to build. Code for building it. Plan mode when the building needs its own mini-plan.
[Section: rule-written]
Rule 02
## If it's not written down, it's not real.
The single biggest shift that separates people who compound with Claude from people who don't is this: they have Claude constantly writing markdown files . Meeting notes. Decisions. Research. Specs. Project status. Their beliefs. Their goals. Everything that normally lives as ephemeral thought-gas becomes a file.
Not because they love filing systems. Because context that's written down is context Claude can use tomorrow. A decision you made in a chat two weeks ago is invisible to the next chat unless it's in a file. A research rabbit-hole you went down last month is lost unless Claude wrote the summary to disk. Your preferences, constraints, and working style are useless to tomorrow's Claude if they only live in your head.
## What to have Claude write down, starting today
- Decisions. When you settle something (pricing, a naming choice, a positioning call), ask Claude to append it to a decisions.md with the date and the reasoning.
- Meeting summaries. Paste the transcript or your notes, ask for a one-page markdown summary with action items. Save to a folder. Never lose a meeting again.
- Research. Every time you go down a rabbit hole, have Claude produce the written summary at the end. That's the thing that's reusable. The chat itself is disposable.
- Who you are and how you work. Your role, your goals, your no-gos, your writing voice. One file Claude reads before every session. This alone changes what you can expect from it.
The medium matters. Markdown specifically is the format where LLMs and humans meet most cleanly. LLMs parse it natively without losing structure, and humans read it at a glance the same way they read a well-written email. Headings, bullets, bold, links. It's portable (any editor, any OS), version-controllable (free history in Google Docs, Notion, or git), and it works across every Claude surface.
Team low-effort start: your team already uses Notion (or a comparable wiki). Make one page (call it "Team AI Context" ) and put your company overview, team roles, active projects, and any shared rules for AI output (voice, banned phrases, data-sensitivity guardrails) on it. Then establish one rule: every Claude chat starts by reading that page first. Some Claude surfaces integrate directly (the Notion MCP for Desktop and Code), but even a manual "paste the URL at the start of each chat" behavior changes what you get back.
Copy the personal seeding prompt to paste into your chats
Paste this once at the top of any long chat (or save it as a snippet with !!me ). Fill in the bracketed fields. You only have to do it once per chat; Claude will carry the context forward for the rest of the session.
Here's context about me and how I work. Please keep this in mind for the rest of our conversation.
**About me**
- Role: [your role]
- Company: [company name + one line on what it does]
- I work primarily on: [what fills your day]
**How I communicate**
- Tone: [direct / detailed / casual / formal]
- Languages: [English / mixed with German / etc.]
- Length preference: [short / thorough / depends on task]
**What I reward**
- [e.g., concrete recommendations with tradeoffs]
- [e.g., admitting when you don't know]
- [e.g., pushback when my logic has holes]
**What frustrates me**
- [e.g., generic advice that doesn't engage with my specifics]
- [e.g., over-hedging, "it depends" answers]
- [e.g., emojis, corporate speak]
**Where to find more context**
- Team shared wiki: [Notion link to Team AI Context page]
- My current projects: [short list or link]
Always check the team wiki link above before answering anything that involves company-specific info. Then help me with: [your actual task].
Copy prompt
[Section: rule-colleague]
Rule 03
## Treat Claude as a colleague, not a mind-reader.
The people who get mediocre results from Claude brief it the way they'd brief Siri: a vague sentence, maximum five seconds of thought, and then disappointment when the output is generic. The people who get great results brief it the way they'd brief a smart colleague who joined the team this morning: enough context to do a good job, clear outputs, stated constraints, one good example.
Claude is not omniscient. It can't read your intent. It has no idea what "done" looks like for you unless you say so. It doesn't know your company's tone, your prospect's industry, your ten previous attempts at the same task. The ceiling on Claude's output is the quality of your brief.
Mind-reader mode
"Write me a follow-up email to this prospect."
Colleague mode
Goal: book a 30-min second call within two weeks. Input: call transcript below. Output: one email, 80–120 words, three paragraphs. Tone: warm but not over-friendly. References two specific things from the call. Avoid: generic "circling back" openers, pricing talk, emojis.
Same model. Ten seconds more effort. The second one ships; the first one gets rewritten. If you've written a good brief once, save it as a skill (see below) and never write it again.
## Colleagues learn new things. Express-onboard them.
A real colleague doesn't pretend to know everything on day one. They ask to read the doc, check the deck, skim the playbook, then come back up to speed. Claude is the same. If there's something Claude needs to know (a new product launch, a prospect's company site, a long internal doc, a PDF that just went out), paste the link or upload the file and say "review this, then we work."
It takes twenty seconds. It saves the next hour of half-informed output. Think of it as express-onboarding a new hire. Same mental model, faster ramp.
The three-times rule. If you've briefed Claude on the same kind of task three times, the brief itself is a skill. The first time you do it, write it. The third time, file it.
Section 02
Two tools worth knowing.
Claude ships as multiple products that solve different problems. Picking the right one for the task is half the battle.
[Section: tool-desktop-vs-code]
Tool 01
## Claude Desktop vs. Claude Code.
Same underlying model. Completely different surface. Both are useful; they're for different moments of your day.
Conversational surface
## Claude Desktop (and Claude.ai / Cowork)
The chat app. Best for thinking out loud, planning, research, drafting, anything that's more dialogue than execution. Non-engineers feel at home here because it's just a chat. This is the right starting point for most BD, marketing, ops, and strategy workflows.
Reach for it when: you want to talk through a problem, draft something with iteration, or onboard yourself to a new topic.
Execution surface
## Claude Code
A terminal-based agent. It can read and write files on your machine, run shell commands, call APIs, use skills, orchestrate subagents. More powerful, more opinionated, more execution-oriented. Requires a bit of setup but unlocks everything in the "Tips" section below.
Reach for it when: the task involves real files, repeated workflows, skills, or anything you want Claude to actually do rather than just discuss.
The honest recommendation for someone starting out: live in Desktop for a week. Get comfortable with long chats and uploaded files. Then, the first time you think "I do this every day, I wish it happened automatically." That's when Code earns its keep.
[Section: tool-chrome]
Tool 02
## The Claude Chrome extension.
When you want Claude to do something inside your browser (navigate to a page, fill a form, pull info from a tab you're already logged into, take a screenshot), you reach for the Claude Chrome extension (https://claude.com/claude-for-chrome). It's Claude with access to your actual browsing session, including your logged-in state on Gmail, LinkedIn, Notion, and whatever else.
Heads up: it burns tokens fast. Every screenshot the agent takes and every DOM snapshot it reads costs context. A single "go to LinkedIn, find this person, summarize their last five posts" task can consume ten to twenty times more tokens than the equivalent conversation in Desktop. Use it for tasks where browser access is genuinely the unlock; don't default to it for things you could paste into a chat.
A small habit that makes it cleaner: keep Chrome dedicated to Claude . Use Arc (https://arc.net/), Safari, Firefox, or a separate Chrome profile for your actual daily browsing, and leave the default Chrome profile empty except for the sites you actually want Claude to touch: Gmail, LinkedIn, your CRM, your Notion. This way the extension's agent-mode has a clean surface with only the tabs it needs, and your personal browsing stays separate. I use Arc for everything and Chrome is basically a "Claude's browser." It stays empty until I want Claude to do something in it.
One safety note: the extension asks before taking destructive actions (sending, deleting, posting), but it can still do things you'd rather it not. Run it with supervision, not in the background, until you trust its patterns on your specific workflows.
Section 03
Tips to compound.
The practical levers that turn "I use Claude sometimes" into "Claude is part of how I work." Pick any one and start this week.
[Section: tip-skills]
Tip 01
## Skills.
A skill is a reusable brief you save once and Claude picks up automatically when the task matches. Technically it's a folder (with a markdown file at minimum, optionally scripts and assets) that Claude Code reads when its description matches what you're trying to do. The practical effect: you stop re-explaining the same task every time.
## A real skill. The first few lines of one of mine.
This is the header of a skill I use every day. It files external sources (URLs, PDFs, files) into my personal knowledge base. You're looking at the minimum viable anatomy: frontmatter, a short body, dispatch rules. The rest of the file is ~150 more lines of mode-specific steps.
---
name: brain-ingest
description: "Ingest an external source (URL, file, folder, or inbox item)
into Marko's personal knowledge base at ~/Documents/Obsidian/brain/.
Use when the user says anything like 'ingest this', 'save this to the
brain', 'remember this article', 'add this paper to my notes', 'process
my inbox', pastes a URL or file path for filing, or asks to add a new
project to the brain."
argument-hint: "<url> | <file-path> | <dir-path> | inbox | (empty for menu)"
user-invocable: true
model: sonnet
allowed-tools: [Read, Write, Edit, Glob, Grep, Bash, Agent, ToolSearch,
mcp__tavily__tavily_extract]
---
# Brain Ingest
Ingest an external source into ~/Documents/Obsidian/brain/.
## Bootstrap (every invocation, no exceptions)
1. Read ~/Documents/Obsidian/brain/CLAUDE.md (the schema)
2. Read ~/Documents/Obsidian/brain/index.md (the catalog)
3. Parse the argument and any flags (--fast, --draft)
## Dispatch by source type
| Argument shape | Mode |
|-------------------|-------------------------------|
| starts with http | URL mode |
| is a file path | File mode |
| is a directory | Directory mode (auto-draft) |
| "inbox" | Inbox mode |
| empty | Menu mode |
...
The description field is the most important part. It's not a summary for humans. It's what Claude scans to decide whether this skill applies to the current task. Make it concrete: "when the user asks for X", "when Y is mentioned", or list the exact phrases. Vague descriptions mean Claude won't trigger the skill when it should.
## Writing tips
- Don't state the obvious. Claude already knows most of what it knows. A skill earns its keep by pushing it toward non-default behavior: your org's preferences, your specific gotchas, the edge cases you keep hitting.
- Build a gotchas section over time. Every time the skill fails in a specific way, add the lesson. The "Gotchas" section is the highest-signal part of any mature skill.
- Don't railroad. Give Claude the information it needs, not a rigid step-by-step script. Skills are reused across slightly different situations; over-specification makes them brittle.
- The folder is the feature. Your skill can include script files, reference markdown, templates, example outputs. Claude reads them when it needs them. The best skills are folders, not just one file.
## Anthropic's public skills. Good starters.
Anthropic publishes a curated set of skills at github.com/anthropics/skills (https://github.com/anthropics/skills). A few worth installing on day one:
skill-creator
The meta-skill. Helps you write new skills correctly. Start here.
frontend-design
Gives Claude opinionated design taste. Avoids the generic "purple gradient + Inter font" AI aesthetic.
docx / pdf / pptx / xlsx
Proper Office document creation and parsing. Essential if your work involves real deliverables.
brand-guidelines
For teams with a defined brand. Keeps Claude's output on-voice.
mcp-builder
When you're ready to wire Claude to a new data source or service.
webapp-testing
Drives a browser to verify your web app actually works. Valuable even for non-engineers shipping marketing pages.
Where your team's skill shelf lives: one shared Google Doc or Notion page is the pragmatic default for a non-engineering team (each skill is a heading, the markdown body below). A shared Drive folder or a Google Chat space pinned with new skills both work as alternatives. Upgrade to a git repo when engineering's ready to help.
[Section: tip-memory]
Tip 02
## CLAUDE.md and agentic memory.
Claude has no persistent memory by default. Every new chat starts fresh. CLAUDE.md is the file that fixes this for Claude Code. It loads automatically at the start of every session, so Claude reads it before it reads anything else in the folder.
The important thing to understand is what CLAUDE.md is for . It's a character sheet, not a manual . It describes identity (role, voice, what this agent optimizes for), locked decisions (things that should never change across sessions), and then a thin index of other files Claude can reach for on demand. Target length: about 100 lines. If yours is over 200, you've put the equipment manual where the character sheet belongs. Move reference material out into sibling files and leave a pointer.
## Don't write it yourself
Inside any folder, run /init in Claude Code. Claude interviews you and writes the first CLAUDE.md for you. From there, make it a habit: at the end of every working session, ask Claude to update the CLAUDE.md and a progress.md with what just changed. The CLAUDE.md absorbs new locked decisions; the progress.md is the rolling diary of what's been shipped. This is how your project's brain stays current without you nagging it.
For Claude Desktop and Claude.ai users, the equivalent lives in Projects : create a project, attach files that carry context (about-me.md, brand-voice.md, current-projects.md, team-wiki.md), and every chat inside that project inherits them.
How agentic memory layers load on top of each other.
## What to put in your CLAUDE.md
- Identity. Role, what this agent is for, voice, what it optimizes for. Top 10–20 lines.
- Locked decisions. Tech stack, naming conventions, no-gos, positioning calls that shouldn't change across sessions.
- Your rules. Short, actionable. "Always quote paths with spaces." "No emojis in code." "Default to German for user-facing copy."
- Thin file index. One line per sibling file (DECISIONS.md, ARCHITECTURE.md, progress.md). Pointers, not contents.
## What NOT to put in it
- Your whole personal biography. That belongs in a team-level context file or a personal brain folder, not in every project's CLAUDE.md.
- Long tutorials or tool manuals. Put those in sibling files (TOOLING.md, SETUP.md) and reference them.
- Week-over-week state. That belongs in progress.md or similar, so CLAUDE.md stays stable.
- Brain content. If you have a personal knowledge base (see below), reference it from CLAUDE.md. Don't copy pages into every project.
Advanced move: give Claude a brain
Andrej Karpathy published a short note on a personal "brain" pattern (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) that's worth reading. The idea: create a folder ( brain/ or memory/ ) with markdown files tracking people you've met, decisions you've made, projects in flight, beliefs you hold, open questions. Reference that folder from your CLAUDE.md so every new chat can pull from it on demand.
The compounding effect is real. Every week the brain has more. Every chat starts from a richer context than the last. What was "I should remember this" becomes "Claude already knows." It's the single highest-leverage piece of infrastructure you can build for yourself once you've been using Claude seriously for a few weeks.
Not everyone needs this on day one. But the first time you notice yourself re-typing the same context across three different projects. That's the signal. Consolidate it into a brain, point your CLAUDE.md files at it, and stop repeating yourself.
[Section: tip-permissions]
Tip 03
## Permissions.
Claude Code asks for permission before running tools: reading files, writing files, running shell commands, fetching URLs, calling an MCP server. For new users this feels relentless: you're approving "read this file" twenty times per session and starting to lose the plot of what you were doing. That's safety by default, but it's not the best long-term mode for you .
The fix: teach Claude which things you've already blanket-approved for a folder or a tool. Permissions live in a settings file: ~/.claude/settings.json globally, or .claude/settings.json inside any project. You don't have to edit it by hand. Ask Claude to do it for you.
Try saying this once: "Add a permission to my global settings that auto-approves all Read, Grep, and Glob calls anywhere, plus read-only Bash commands like ls, cat, head, tail, pwd, git status, git diff, git log. Keep everything else asking."
Claude will make the edit and explain what it did. From that point on, the obvious stuff stops interrupting you.
## Safe to auto-approve vs. keep asking
Safe to auto-approve
- Read-only file tools: Read , Grep , Glob , Ls
- Read-only shell: ls , cat , head , tail , pwd , git status , git diff , git log
- Public fetches: Tavily search, web docs
Keep asking
- File writes: Edit , Write , NotebookEdit
- Mutating shell: anything that changes files or state
- Destructive commands: rm , git push --force , dropping tables, git reset --hard
- Anything touching shared infrastructure or production
You can also scope permissions per-folder: "always allow Bash in this project but ask everywhere else." Rule of thumb: relax permissions where the blast radius is small; keep them tight where mistakes cost you. Your own sandbox folder is small-radius. A shared CRM or production database is not.
[Section: tip-session]
Tip 04
## Session economics.
Every conversation with Claude burns tokens, and every plan has a cap. On the $20/month tier you will hit session limits. Fast, if you let a single chat sprawl for hours. These are the moves that stretch a subscription.
## Pick the right model for the task
Claude ships in tiers. Haiku is fastest and cheapest, good for simple drafts, summaries, short questions, and anything where you'd happily accept "good enough." Sonnet is the workhorse, your default for most real work. Opus is the heaviest model. Reserve it for genuinely hard reasoning across a lot of context, or when Sonnet has already failed. Using Opus to ask "what day is it" is throwing tokens away.
## Dial thinking intensity
Claude Code has thinking budgets you can ask for by name. think is light. think hard is more. ultrathink is maximum. Every level burns more tokens. Most tasks don't need ultrathink. It's for multi-step reasoning, architectural decisions, or debugging something slippery. For writing a memo, drafting an email, or summarizing a doc, default thinking is plenty.
## Use subagents for heavy lifting
When a task involves reading a lot of material (scraping a site, digesting a big PDF, running a broad search across your codebase), dispatch a subagent instead of doing it in your main chat. The subagent burns its own tokens, does the work, and returns a summary. Your main session only sees the result, not the forty pages of raw material. This is how Anthropic's own engineers manage their context budgets, and it's the single biggest move for keeping long sessions efficient.
## Compact or clear long sessions
Every turn of a conversation re-sends the prior context. Long sessions get expensive fast. Use /compact to summarize the conversation so far (Claude distills it and continues with the distilled version), or /clear to start fresh with only what matters. Doing this proactively (before you hit the wall) is much cheaper than being forced to by the limit.
## Be specific upfront
A vague prompt leads to back-and-forth clarification, which burns tokens on every turn. Three rounds of "actually what I meant was…" costs more than a clear brief in the first message. This is Rule 3 applied to your wallet.
A good mental model: treat your monthly session budget the way a freelance consultant treats their 40 billable hours a week. You don't burn them on low-value meetings that could've been emails. Pick the right tool, pick the right model, pick the right depth. Most people who "run out of Claude" aren't using too much. They're using it too shallowly.
[Section: first-week]
After the workshop
## Your first week.
Pick one thing. Ship it. Tell your team what you picked.
## By Sunday night
Ship one skill. One of your repeat tasks. Thirty minutes to write. Save it into your team's shared Notion page.
## By next Friday
The team has five. Share yours. Steal theirs. Your team's shelf is real.
## If you get stuck
- If you can't pick a task, pick the one you'd most hate explaining to a new hire on their first day. That's the one.
- If the skill is bad, use it anyway. Fix tomorrow's version.
- If you don't have time, write a five-line version on your phone during a meeting you shouldn't have been in. That's the MVP.
- If you're not sure it's working, show it to one teammate. If they laugh and then ask for a copy, you shipped something.
## Questions after the workshop?
Email me directly at marko.stokic@proton.me (mailto:marko.stokic@proton.me). Tell me what you built, what broke, what the team stole first.
If you want the full workshop for a different team, or a half-day diagnostic for your org, the intake form is here (https://stokic.ai/intake).
[Section: resources]
Resources
## Links worth bookmarking.
People and pages I found genuinely useful while putting this together. Drop in when you want to go deeper on any of it.
CC for Everyone
An interactive Claude Code guide that runs inside Claude Code itself. You learn the tool by using it. Best starting point if you are completely new.
ccforeveryone.com → (https://ccforeveryone.com/)
Anthropic's public skills library
The curated set of official skills: skill-creator, frontend-design, docx/pdf/pptx/xlsx, mcp-builder, and more.
github.com/anthropics/skills → (https://github.com/anthropics/skills)
Skills docs: Claude Code
Official docs on frontmatter fields, configuration options, hooks, and distribution.
code.claude.com/docs/en/skills → (https://code.claude.com/docs/en/skills)
Lessons from Building Claude Code: How We Use Skills
Anthropic engineering post (March 2026), cataloguing skill types, gotchas, progressive disclosure, distribution. The single best read on skills in practice.
x.com/trq212 → (https://x.com/trq212/status/2033949937936085378)
Ultimate Claude Cowork Starter Pack
Corey Ganim (March 2026), plugins, context files, workflows. Light on depth but useful for seeing the shape of a mature Cowork setup.
x.com/coreyganim → (https://x.com/coreyganim/status/2033206539595461070)
Andrej Karpathy's personal brain pattern
Short note on structuring a personal knowledge base Claude can read from. Inspiration for the "brain/" pattern in Tip 02.
gist.github.com/karpathy → (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f)
Claude for Chrome
The extension page itself: install, permissions, safety defaults.
claude.com/claude-for-chrome → (https://claude.com/claude-for-chrome)
stokic.ai (https://stokic.ai/) ·
beliefs (https://stokic.ai/beliefs) ·
publications (https://stokic.ai/publications) ·
speaking (https://stokic.ai/speaker)
Marko Stokić · Munich