Claude Cowork
Workshop · Day 2 · Morning

Skills Design
& Your Second Brain

From raw skills to intelligent architecture — design, organise, and connect.

Skills Design
Obsidian
MCP Demo
~90 min
01 / 29
Claude Cowork
Recap · Day 1

Where we were. Where we're going.

Yesterday
What are skills?
The three-tool model. Skill vs reference vs plugin. Built your first skill and ran it inside Claude Cowork.
Today
How do we design them well?
Monolithic vs granular. Where skills live. Connecting them to your knowledge base.
Next milestone
A connected second brain
Your skills + your knowledge + Claude — all wired together via MCP.
This morning · 90 min
1
Skills Design
30 min
2
Obsidian
30 min
3
MCP Demo
15 min
4
Q&A
15 min
02 / 29
Claude Cowork
The Big Question

Before you build, decide three things.

Most people get skill architecture wrong on the first try — going too big, too granular, or by the wrong axis. These three questions force the call early.

?
Question 01 · Size
Big skill
or small skills?
One file that does the whole job — or many focused pieces wired together?
?
Question 02 · Axis
Role-based
or task-based?
Organise by who is acting, by what is done, or by what gets produced?
?
Question 03 · Scale
One agent
or many?
A single skill that runs solo, or a pack of specialised agents collaborating?
By the end of this module, you'll have a framework for answering all three before writing a single line.
03 / 29
Claude Cowork
The Skill Spectrum

From one giant file to a constellation.

MONO-
LITHIC
MIDDLE
GROUND
GRAN-
ULAR
1 FILE · ALL STEPS
N FILES · 1 STEP EACH
M
Monolithic
Fast to write — one file, one place
→ Easy to read top-to-bottom
→ Hard to reuse parts elsewhere
→ Hard to debug when one step fails
G
Granular
Reusable across workflows
→ Different agents can pick & mix
→ Debug one piece without breaking others
→ Higher upfront design cost
Neither extreme is right by default. Where you land depends on reuse, complexity, and how many agents will touch it.
04 / 29
Claude Cowork
Demo · Monolithic

One file. All the steps.

An expense-claim skill written as a single monolithic file. It works — until you try to reuse, debug, or scale it.

expense-claim-skill.md
# Expense Claim Skill

## Steps
1. Read receipts
2. Validate amounts
3. Match to budget codes
4. Route to approver
5. Send notification
6. Generate PDF
7. Upload to Drive
Problems
1
Hard to reuse
Receipt parsing logic is locked inside this skill. Can't borrow it for invoice processing.
2
Hard to debug
PDF wrong? You re-run all seven steps to test the fix. Slow feedback loop.
3
Hard to hand off
A different agent can't pick up the "approval-router" piece on its own.
05 / 29
Claude Cowork
Demo · Granular

Same job. Five reusable pieces.

📁 skills/finance/
├── receipt-intake.md → parses
├── budget-validator.md → checks codes
├── approval-router.md → routes
├── notification-sender.md → alerts
└── pdf-generator.md → outputs
5 SKILLS 1 RESPONSIBILITY EACH FULLY REUSABLE
Benefits
Reusable across workflows
receipt-intake also powers invoice claims, mileage, petty cash.
Agent-friendly
CFO Agent uses validator + router. Admin Agent uses sender + generator. Different agents, shared pieces.
Modular debugging
PDF wrong? Test just pdf-generator.md. Other steps untouched.
06 / 29
Claude Cowork
Decision Framework

Four questions. Two paths.

Ask yourself…
If YES →
If NO →
1. Is this skill reusable across different contexts?
Break it out
Keep together
2. Does it have clear decision branches?
Specialise it
One step
3. Could multiple agents use a piece of it?
Modularise
Monolithic
4. Does it need a different persona or tone?
Separate skill
Same skill
Rule of
thumb
Role-based
Who is acting · CFO Agent, Admin Agent
Task-based
What is done · Receipt Parser, Letter Drafter
Outcome-based
What is produced · Expense Report, Proposal PDF
07 / 29
Claude Cowork
Hands-On · 10 min

Your turn. Pick a workflow.

Take a real workflow from your daily work. Map it monolithically first, then break it apart.

1
Map monolithic
Write one big skill file. List every step from start to finish. Don't optimise — dump it all.
2
Split into 3–5
Apply the 4-question framework. Break the monolith into 3–5 micro-skills. Name each one.
3
Pick your axis
Role-based, task-based, or outcome-based? Share back with the group: what did you choose and why?
10 minutes · solo work · then 5-minute share-back. Pick a workflow you already do weekly — that's where reuse pays off.
08 / 29
Claude Cowork
Section 2 · Transition

Now your skills
need a home.

A place where your skills, your research, your client intel, your agent outputs — all of it connects together. Introducing your second brain.

09 / 29
Claude Cowork
What is Obsidian?

Not a notes app. A network.

Obsidian is a markdown editor that turns your notes into a connected knowledge network — every note is a node, every link is an edge.

100% local markdown files
Yours forever. No lock-in. Sync however you want.
Bracket-linking [[like this]]
Connections, backlinks, graph view — emergent structure.
MCP-ready
Claude can read, search, and write into your vault.
Graph view · simulated
Knowledge becomes a network
hub
10 / 29
Claude Cowork
The Problem

Why your knowledge keeps dying.

Today · the chaos
📧Important context in email threads
💬Decisions buried in Slack
📑Specs scattered across Google Docs
🧠Critical knowledge in one person's head
📝Notes apps that never get re-opened
Consolidate
into vault
After · one place
Everything searchable, instantly
Connections between ideas surface automatically
Knowledge survives when people leave
Claude can read it all via MCP
Agents can write back into the vault
11 / 29
Claude Cowork
Three Layers

Start flat. Layer up.

A knowledge base evolves in three stages. Don't skip — each layer unlocks the next.

01
Flat Notes
Markdown files in folders. Use tags for basic organisation. No linking yet — just dump and structure.
START TODAY
02
Linked Notes
Use [[bracket linking]] to connect ideas. Backlinks reveal hidden relationships. Your knowledge becomes a network, not a list.
THIS WEEK
03
AI-Connected Brain
MCP connects Claude to your vault. Agents can read, write, and query your knowledge base. Your second brain becomes your team's intelligence layer.
NEXT MILESTONE
You don't need layer 3 on day one. Start with 5 flat files. Add links when it feels natural. Add MCP when the vault is worth querying.
12 / 29
Claude Cowork
Live Demo

Marcus's Digital Marketing OS.

An Obsidian vault built over 18 months. 100 markdown files. 21 folders. Hundreds of backlinks. Queryable by Claude.

100
Markdown files
From client briefs to campaign learnings.
21
Folders
Projects, resources, areas, archive.
1
Queryable brain
Claude reads it all via MCP.
DEMO 01
Folder tour
DEMO 02
Live search
DEMO 03
Backlinks
DEMO 04
Graph view
I didn't build this overnight. I started with 3 files. This is where you're going — not where you start.
13 / 29
Claude Cowork
Starter Structure · PARA

Five folders. Start here.

📁 My Second Brain/
├── 📁 00 - Inbox/
├── 📁 01 - Projects/
├── 📁 02 - Resources/
├── 📁 03 - Areas/
└── 📁 04 - Archive/
00
Inbox
Dump everything here first. Sort later, not now.
01
Projects
Active work with a deadline — campaigns, builds, launches.
02
Resources
Reference material — guides, templates, research, prompts.
03
Areas
Ongoing responsibilities — health, finance, team management.
04
Archive
Completed or inactive — not deleted, just out of the way.
14 / 29
Claude Cowork
Markdown Cheat Sheet

Plain text. Beautiful output.

Markdown is the universal language of notes. Learn these eight patterns and you can write any Obsidian note fluently.

# Heading 1
Big title
## Heading 2
Section title
**bold text**
bold text
*italic text*
italic text
- bullet point
• bullet point
1. numbered
1. numbered
[[Link to note]]
→ creates a backlink
#tag
→ filterable label
The killer feature
Type [[ in any note and Obsidian autocompletes from your vault. Every link creates a two-way connection — the linked note knows it's referenced.
15 / 29
Claude Cowork
Hands-On · 15 min

Build your starter vault.

1
Download
Obsidian from obsidian.md (free)
2
New vault
Name it My Second Brain
3
Folders
Create the 5 PARA folders
4
3 notes
On a topic you already know well
5
Link them
Use [[note name]]
6
Graph view
Open it · see your network emerge
By end of 15 min
Working vault · PARA structure · 3 notes linked together · graph view showing the network. That's already a second brain.
16 / 29
Claude Cowork
Going Deeper · Brain OS

PARA tells you where.
Brain OS tells you what to do.

The starter vault is a folder system. The Brain OS is an operating system for thinking — built on three philosophies that compound across years.

310
total files
57
wiki articles
102
OS docs
2am
nightly compile
Marcus's live vault · May 2026
17 / 29
Claude Cowork
Two Philosophies

The thinking method + the writing method.

Philosophy 01
Zettelkasten
Niklas Luhmann · 70 books, 400 articles
01One idea per note — atomic, not a topic dump
02Unique ID — permanent, links never break
03Link liberally — ideas build on ideas
04Write in your own words — never copy-paste
05No rigid hierarchy — structure emerges from links
A Zettelkasten is not a filing system — it's a thinking partner.
Philosophy 02
Karpathy-Style
Andrej Karpathy · compilation workflow
The fundamental shift:
human curates · LLM writes.

Most knowledge systems collapse because maintaining them is too laborious. You capture 100 articles and never process them.

Brain OS fix
Drop raw sources into the vault. Claude compiles them into structured wiki articles. Indexes stay current automatically.
18 / 29
Claude Cowork
Vault Anatomy · 7 Layers

The vault mirrors how your brain works.

Folder structure
📁 2ndBrain/
├── 📁 inbox/ ← drop zone
├── 📁 raw/ ← immutable sources
├── 📁 wiki/ ← compiled knowledge
├── 📁 os/ ← how-to-do-things
├── 📁 areas/ ← ongoing life
├── 📁 projects/ ← active focus
└── 📁 outputs/ ← what gets made
Brain analogue
inbox
Working memory — new captures, unprocessed
raw
Sensory input — sources, never edited
wiki
Semantic memory — compiled evergreen knowledge
os
Procedural memory — SOPs, playbooks, brand
areas
Episodic memory — journals, meetings, tasks
projects
Active focus — deadline-driven work
outputs
Synthesised creations — answers, slides, posts
19 / 29
Claude Cowork
PARA × Brain OS

Same folders. Bigger job.

Brain OS extends Tiago Forte's PARA framework with one critical addition — a compilation step that transforms saved resources into living wiki articles.

PARA layer
Brain OS equivalent
What happens to it
Projects
projects/
Active work with deadlines
Areas
areas/
Ongoing responsibilities — journals, meetings
Resources
raw/ wiki/
Compiled into permanent knowledge
Archives
archive/
Completed, searchable history
PARA says
Where to put things
Brain OS adds
What to do with them
20 / 29
Claude Cowork
The Provenance Chain

Every output traces back to a real source.

Three layers of memory. Every claim in a wiki article links back to the raw file it came from. Every output links back to the wiki article it drew on.

LAYER 01 · RAW
Raw source
What I read · article, transcript, paper, meeting note
raw/article-x.md
compile
LAYER 02 · WIKI
Wiki article
What I understood · synthesised, atomic, evergreen
wiki/concepts/idea.md
synthesise
LAYER 03 · OUTPUT
Output
What I made · answer, slide, post, proposal
outputs/qa/2026-05-24.md
Every output → wiki article → raw source. No claim is hallucinated.
This is what separates genuine thought leadership from AI noise. Provenance compounds trust.
21 / 29
Claude Cowork
Core Workflows

Four loops that run the system.

INGEST
Drop into raw/. Claude registers, reads, identifies touch-points, compiles wiki articles.
COMPILE
One concept per file. YAML frontmatter, dense cross-links, every claim traces to a raw source.
QUERY
Read index → search articles → draft cited answer → save to outputs/qa/.
DREAM
Nightly at 2am · dream.py via launchd. Compiles backlog, health-checks vault, writes log.
The compound result · live scale
310
total files
83
raw sources
57
wiki articles
102
OS docs
40+
concepts
9
company pages
queryable now
The vault grows while you sleep. This is what compounds.
22 / 29
Claude Cowork
Model Context Protocol

The bridge between Claude and your vault.

MCP — Model Context Protocol — is the open standard that lets Claude read, search, and write to external systems. Including your Obsidian vault.

Claude
Reasons · drafts · queries
PROTOCOL
MCP
read · write · search
Your Vault
100 notes · 21 folders
READ · search your notes mid-conversation
WRITE · agents append findings into the vault
QUERY · ask in natural language, get sourced answers
23 / 29
Claude Cowork
Use Case 01 · Personal Query

Ask Claude. It answers from your vault.

claude.ai · MCP connected
vault online
What did we learn from the Q3 retainer pitch last year?
→ searching vault for "Q3 retainer"…
→ reading 02-Resources/pitches/q3-retainer-debrief.md
→ cross-ref [[Client Background — ACME]]
From your debrief: 3 lessons — pitch deck was too dense (cut by 40%), the CFO's main objection was implementation risk, and the relationship lead at ACME (Sarah Tan) is now at Wavelength as VP Marketing.
What happened
Claude turned your scattered notes into a sourced answer in 2 seconds.
No copy-pasting old docs into the prompt
Cites the exact file paths it pulled from
Connects across notes you'd forgotten linked together
SCALE · PERSONAL
24 / 29
Claude Cowork
Use Case 02 · Agent → Vault

Your agents grow your vault.

SOURCE
Live meeting
Granola, Otter, transcript
AGENT
Summary skill
extract · structure · link
VAULT
New note created
2026-05-24-acme.md
# Meeting · ACME Pitch Review · 24 May 2026
## Decisions
- Drop slide 7 (too data-heavy)
- Sarah owns next-round prep
## Linked
[[Client — ACME]] · [[Q3 Pitch Deck]] · [[Sarah Tan]]
#meeting #project/acme #pitch
Your vault grows without you. Every meeting, every research session, every customer call — captured and linked, automatically.
25 / 29
Claude Cowork
Use Case 03 · Company Brain

One vault. Whole team.

SHARED VAULT CEO CFO OPS SALES PM DEV
6 contributors · 1 brain · ∞ queries

Each team member writes into the same vault. Claude queries it for any team member — so the answer is always shaped by everyone's contributions, not just yours.

→ CEO asks · gets context from CFO's notes
→ New hire onboards · reads 18 months of decisions
→ Someone leaves · the knowledge stays
SCALE · COMPANY
26 / 29
Claude Cowork
What You Can Build

Four scale levels. Same stack.

The same Obsidian + MCP combo scales from a personal journal to enterprise-wide intelligence. Pick the level that matches your stage.

LEVEL 01
Personal brain
Your notes, your queries, your reasoning partner.
LEVEL 02
Team base
Shared notes across 3–10 people. Onboarding solved.
LEVEL 03
Company layer
Org-wide intelligence. Survives turnover. Compounds.
LEVEL 04
Agent memory
Agents read, write, and reference vault as long-term memory.
Compounding
Each level builds on the last. Personal habits → team rituals → company architecture → autonomous agents. You don't pick a level — you climb one.
27 / 29
Claude Cowork
Tonight's Homework

Three things. One week.

1
TONIGHT
Download Obsidian
Grab it from obsidian.md. Build the 5 PARA folders. Create your first 3 notes — link them together.
2
THIS WEEK
Reach 10 linked notes
Add a note per day. Link aggressively. Open Graph View by Friday — see the network you've built.
3
NEXT SESSION
Wire up MCP
We'll connect Claude to your vault together. Come with your 10 notes and a question you wish Claude could answer.
Materials to grab
Obsidian app · obsidian.md
PARA folder template
Markdown cheat sheet
Sample skill files
28 / 29
Claude Cowork
Day 2 · Complete

Q&A & what's next.

TAKEAWAY 01
Design skills modular — not monolithic by default.
TAKEAWAY 02
3 files linked together is already a second brain.
TAKEAWAY 03
Scale at your pace — personal, team, company, agent.

Floor's open. Questions, edge cases, war stories — bring them.

marcus@aitraining2u.com
aitraining2u.com
29 / 29