Personal Talk · Chris Erler

My Setup.

From a personal AI twin to organizational brains. One pattern, two scales. The hours saved, the hires never made, and the master template you fork tonight.

Christoph ErlerEO AI Exchange #1 · May 11, 2026

EO AI Exchange #1
01 / 51
Where this comes from

The bet I made.Build infrastructure, not a team.

After ComX, I had the option to hire and rebuild what I already knew. Instead I bet that AI as infrastructure can replace the team I would have hired. Three years later, that bet is still the strategy.
Three rules I follow now.
Every workflow I touch becomes a skill. Every meeting I have becomes a memory. Every recurring task lands on a cron.
Zero employees. Just a brain, version-controlled and queryable. Runs at 150 EUR per month.
Freedom is presence of choice, not absence of work.
EO AI Exchange #1
02 / 51
The math behind the bet

1.5 FTE never hired.
One brain doing that work.

150EUR/mo
The brain doing the same work
Claude Pro subscription
Telegram VM on AWS Lightsail Frankfurt
Minor API spend across MCP tools
All-in monthly, two years running, single seat.

Cost ratio: about 45x to 60x. Honest equivalent-capability comparison, not full-time-hires fantasy. Plus zero coordination tax, zero hiring cycle, zero performance review. The bet was a no-brainer once the math was on paper. The rest of tonight is what I do with the saved capacity.

EO AI Exchange #1
03 / 51
Where the hours go

The math of bad setup.

4h
Tool research, signup, abandonThe new shiny that lives 6 days on your dock.
/wk
3h
Re-explaining context to a new chat"Here is who I am, here is what I do, here is the project."
/wk
2h
Patching when an integration breaksWorkspace updates, API deprecations, OAuth re-auth.
/wk
2h
Switching between 5 chat appsClaude here, ChatGPT there, Gemini for Workspace, Perplexity for search.
/wk
2h
Manual prompt copy-paste ritualsThe same 10 prompts you have typed 400 times this year.
/wk
1h
Reading "what is new" instead of using what worksFOMO disguised as professional development.
/wk
14 hours per week, every week, until the architecture replaces the rituals. This is what we are killing tonight.
14h/ wk
EO AI Exchange #1
04 / 51
Act 1 · My Personal Brain

My personal brain, in four numbers.The stack that takes 23 hours of routine off my week. Inbox, calendar, briefings, drafts, real estate pricing, content. All four numbers below, in production for two years.

150EUR / mo
Total stack
All-in: API, infra, voice, SaaS
31
MCP Tools
Across 8 categories, all production
19
Skills
Markdown procedures, judgment + parameters
125
Memory files
1-page profiles, diarized from meetings

Replaces what would otherwise cost 7 to 9k EUR per month in 1.5 FTE of part-time hires (half-time EA + Analyst + Content Manager). Payback in 4 to 5 months including 200h setup time.

EO AI Exchange #1
05 / 51
Act 1 · My Personal Brain in detail
My Personal Brain stack
One operator. Six hats. Zero employees.

Erler Ventures is my holding company. Inside it: my advisory practice (with clients), real estate across 7 properties, plus aviation training and personal life. Same brain runs all six.

The six hats
Erler Ventures (holding) · 2 advisory clients · 7 properties · pilot training · personal life
Skills (my USP)
19 named clusters · the actual tasks I do, codified. The work I refuse to hire away because it IS my edge.
Short-term memory
125 .md files, loaded per task / person / project
Body (tools)
39 Workspace tools orchestrated (Gmail, Calendar, Notion, Holidu, Close, more)
Monthly cost
150 EUR (Claude Pro + Telegram VM + minor APIs)
Hours saved
~23h / week (personal admin + client work), tracked over 8 weeks
What this brain is worth, two angles
1. Equivalent capability:
0.5 EA + 0.5 Junior Analyst + 0.5 Content Manager = 80 to 110k EUR / year (1.5 FTE, DACH all-in)

2. My own time freed:
~23h/week × 47 weeks × 250 EUR/hr = ~270k EUR / year (opportunity cost, hours redirected into building)

The brain costs 1.8k EUR per year. Either lens beats that by 40 to 150x.
B
Brain Bot
cron 07:00 · daily
Sun, 11 May
Top of inbox
1. [redacted PE OP] diagnostic, response due 12:00
2. Parkside Knöringen velocity audit, [host] WhatsApp
3. Diana wedding venue, decision deadline today
Today, 4 calls
09:00 [redacted] strategy review
11:00 Raphi RC sync, pricing pivot
14:00 [redacted] Mala status
16:00 Stephan Cape Town call
Open threads
Hölderlin financing · ehevertrag draft · FOS pricing
07:02
EO AI Exchange #1
06 / 51
A day in my life with the brain

Where my 23 hours per week actually come from.Personal admin and ops + advisory work for clients + real estate ops. Real tasks, real frequencies, measured over 8 weeks.

Task When Without brain With brain · what I actually get Saved / week
Morning briefingtop inbox, calendar today, open threads Daily, 7am cron 30 min coffee + inbox + calendar review 3 min Telegram read, my whole day in one page 2.25h
Meeting prepfull context per person before any call ~4 calls / day 20 min × 4 LinkedIn + past mail + CRM hunt 3 min × 4 dossier ready: who, last contact, open threads, talking points 5.7h
Email + WhatsApp draftsvoice-locked per relationship ~8 drafts / day 5 min × 8 writing in someone else's voice 1 min × 8 draft ready, my tone per person+channel 2.7h
Real estate pricing6 studios via Holidu Daily auto 1h / day manual Holidu review + adjustments 0 min cron runs at 06:00, I read the diff over coffee 5h
Stakeholder updatesquarterly investors, monthly clients, weekly co-founders ~5 / week 20 min × 5 writing the update from scratch 5 min × 5 brain drafts from last contact + open threads 1.25h
Friday weekly reviewacross all 7 hats Friday 17:00 90 min reflection + drafting note to self 10 min review draft, edit, send to memory 1.3h
Invoices + DATEVmonthly receipts batch Monthly auto 2h / month manual DATEV upload 0 min brain forwards, I just approve in email 0.5h
Client diagnosticsprocess audits, ICP analysis, sales script reviews ~2 engagements / month 8h × 2/mo manual deep-dive analysis 1.5h × 2/mo /audit-process or /sales-script-rewriter, I review 3h
Client deliverablesstrategy frameworks, MoMs, action items, briefings ~3 / week 90 min × 3 drafting from scratch 15 min × 3 brain drafts in my voice, I edit and ship 3.75h
Total time saved per week ~23h
Tracked over 8 weeks (March-April 2026). Personal admin + ops + advisory work for clients = ~23h per week reclaimed. Real tasks. Real frequencies. Real outputs.
EO AI Exchange #1
07 / 51
Act 1 · the skills that do the work

Eight skills, forkable tonight.Five universal for any founder. Three domain examples to adapt to your own vertical.

What is a skill?
A recipe Claude reads every time.
One Markdown file. Tells the AI three things:
  • WHEN to fire (trigger word)
  • WHAT to do (steps in plain English)
  • HOW you sound (voice + context)
Write it once. Run it forever. Smallest reusable unit of AI work.
Like a saved macro, but the macro can think.
EO AI Exchange #1
08 / 51
My Personal Brain · how it works

A brain has four parts. So does mine.Same shape as your own. Right column shows what happens on Friday at 17:00 when /weekly-review fires.

Layer 1 · Long-term Memory
Everything I might ever need.
Like your childhood and skills. Always there, never recalled until needed.
Friday 17:00
Cron fires. Archive stays idle. Brain just knows: run /weekly-review.
Layer 2 · Short-term Memory
What I just loaded for this task.
Like the phone number you just looked up. Held just long enough to use it.
What loads
~10 small files: review skill + 6-hats system + current priorities + last week's note. About 5KB total.
Layer 3 · Working Memory
What thinks, right now.
Like your inner voice. Active, in-the-moment reasoning. Stores nothing.
What happens
Walks all 6 hats, ranks open threads, drafts the Friday note. ~90 seconds.
Layer 4 · Body
What I actually do.
Like your hands typing. Predictable, real-world execution via tools.
What I get
1-page brief in Telegram at 17:05. Stale threads archived. 3 decisions flagged for Monday.
Same shape as your own brain. Just persistent and queryable. 23h saved per week. Next slide: how the brain saves what it just learned.
EO AI Exchange #1
09 / 51
Same brain · real components

Same four parts. Real components.Each layer mapped to what actually runs on my machine. Plus the one hard thing about building each.

Layer 1
Long-term Memory
What runs
Notion (truth data) + github.com/chris1928a/erler-brain-v3 + Drive (originals). Three sources, one canonical role each.
Hard thing
Keeping all three consistent. Drift between Notion, GitHub and Memory kills trust within weeks. Weekly /audit-memory catches it.
Layer 2
Short-term Memory
What runs
~/.claude/projects/erler-brain/memory/ with 125 .md files plus a 50-line CLAUDE.md that routes which files load per task.
Hard thing
Picking what loads. Too little = no context. Too much = blown context window + degraded reasoning. Tuning takes weeks.
Layer 3
Working Memory
What runs
Claude Code on an AWS Lightsail VM in Frankfurt. EU residency, audit-logged, read-only by default. Cron triggers via systemd.
Hard thing
Cold-start latency, Anthropic rate limits during peak hours, sandboxing destructive actions, keeping the EU boundary for GDPR.
Layer 4
Body
What runs
39 MCP tools (Gmail, Calendar, Notion, Drive, Holidu, Close, GitHub) + Telegram bot as my interface + 7 cron jobs across the week.
Hard thing
OAuth refresh cycles, Workspace API breaking changes, MCP server crashes. Maintenance tax is ~2h per month if I stay disciplined.
Same body analogy as the previous slide. Just every layer mapped to a real file path, repo, or running process you can audit. The next slide names the six tools that play those four layers.
EO AI Exchange #1
10 / 51
Body-mapped · the six tools

Same brain. Six tools.Each layer played by 1 to 3 tools. Body analogy on the right. Deep dives in the appendix from slide 28 onward, each with a body-association badge.

Layer 1
Long-term Memory
Tools playing this layer
GitHub, code + skills, 3 repos
Notion, shared prose, hand-edited truth data
Drive, originals only (PDFs, decks, contracts)
Body analogy
Like your childhood plus skills plus diploma drawer. Always there, never recalled until needed. 3 tools because the layer is big.
Layer 2
Short-term Memory
Tool playing this layer
Memory files, 125 1-pagers in ~/.claude/projects/erler-brain/memory/, routed by CLAUDE.md (~50 lines).
Body analogy
Like the phone number you just looked up. Held just long enough to use it. 1 tool, on-demand load per task.
Layer 3
Working Memory
Tool playing this layer
Claude Code, thin harness running the model in a loop. ~50-line CLAUDE.md, AWS Lightsail Frankfurt for crons, GDPR-aware.
Body analogy
Like your inner voice. Active, in-the-moment reasoning. Stores nothing. 1 tool as the engine.
Layer 4
Body
Tools playing this layer
Telegram bot, phone interface, voice + text
39 MCP tools, Gmail / Calendar / Notion / Drive / Holidu
7 cron jobs, daily, weekly, monthly autopilot
Body analogy
Like your hands typing and mouth speaking. Predictable, real-world execution. 3 tools because execution has many surfaces.
If you remember nothing else, remember this map. Four layers, six tools, one brain. Next five slides are 30-second snapshots, one per tool. Full deep dives in the appendix from slide 33 onward.
EO AI Exchange #1
11 / 51
Layer 1 · Long-term Memory · Snapshot · Tool 01
Code + Skills
GitHub.
Where my code and skills live in version control. The brain's hard drive.
"Every skill I write is a permanent upgrade. Build it once, runs forever."
Three things to know
3 repos
erler-brain-v3 (private master) + eo-ai-exchange (public) + cape-town-ai-exchange (public)
1.72 GB
~200 .md
17 skill clusters + 125 memory files mirrored from ~/.claude/projects/
git-tracked
never in
OneDrive. Index corruption risk. Repo lives in C:\Code only
anti-pattern

Deep dive: slide 33. Full repo structure, path conventions, anti-patterns I learned the hard way.

EO AI Exchange #1
12 / 51
Layer 1 · Long-term Memory · Snapshot · Tool 02
Shared prose
Notion.
The brain hub for non-developers. Where I share with collaborators and store hand-edited truth data.
"Skills + frameworks live in GitHub. Meeting notes, status updates, deep dives with numbers live in Notion."
Three things to know
9 hubs
Personal, Ventures, Real Estate, Aviation, Content, Innovation, etc.
hand-edited
full content
Never link-rot to "see file X.md". The prose itself lives here, always.
strict rule
audience
Collaborators who do not read code. Skills stay in GitHub, prose comes to Notion
sharing

Deep dive: slide 34. All nine hubs, what lives in each, the no-link-rot rule.

EO AI Exchange #1
13 / 51
Layer 1 · Long-term Memory · Snapshot · Tool 03
Originals only
Drive.
Source-of-truth originals. PDFs, decks, contracts. Not a brain anymore. That was a costly lesson.
"Drive-as-brain failed at API rate limits and 200-2000ms latency. ripgrep returns in <50ms."
Three things to know
lives here
Contracts, decks, legal, aviation charts, DATEV uploads. Read-only originals.
PDFs + PPTX
not here
Skills, memory, runtime data. Drive is too slow + drift-prone for that.
by rule
v4 lesson
Tried Drive-as-brain. 6 weeks of half-built sync. Killed April 27.
painful

Deep dive: slide 35. Folder structure + why I removed Drive from the brain layer entirely.

EO AI Exchange #1
14 / 51
Layer 3 · Working Memory · Snapshot · Tool 05
Thin harness
Claude Code.
The harness that runs the model in a loop. Reads files, manages context, enforces safety.
"Loop, files, context, safety. Four things only. Push intelligence up into Skills, not into the harness."
Three things to know
CLAUDE.md
~50 lines, pointer-only, always loaded. Routes which skill + memory file load per task.
router
search
ripgrep + glob agentic search, no vector-DB. Modern models reason over text directly.
<50ms
MCP
15+ adapters: Notion, Gmail, Calendar, Telegram, Computer-Use. All optional, per-session.
plug-and-play

Deep dive: slide 45. Harness layer in full: CLAUDE.md, sub-agents, hooks, MCP servers.

EO AI Exchange #1
15 / 51
Layer 4 · Body · Snapshot · Tool 06
Mobile interface
Telegram.
My brain on my phone. Voice + text triggers from anywhere. Calls back to Claude API with full brain context.
"I read this on my phone in bed before I move. By the time I am at the desk, I know what today is about."
Three things to know
server
AWS Lightsail Frankfurt. Sovereign EU stack, GDPR-aware, ~20 EUR/month.
24/7
bot
39 tools registered. HMAC webhook verification, hourly Git sync to pull latest skills + memory.
production
daily
morning brief, prep meeting, draft outreach, log this call.
~23h/wk saved

Deep dive: slide 46. Full bot architecture, all daily triggers, voice + text routing.

EO AI Exchange #1
16 / 51
After every session

How I actually save what I learn.Not one memory. Four layers. Each layer has one job. After a session, the brain routes the content to where it belongs. No layer auto-feeds another. Audit skills catch drift.

Layer 1 · Cache
User Memory
~/.claude/.../memory/
What lives here
My own preferences, voice rules per contact, feedback to remember, project status snapshots. Compressed pointers, never the truth itself.
When it goes here
I say it out loud: “remember that ...”, “forget about ...”, “update: X is now Y”. Never auto-written.
Layer 2 · Truth Data
Notion
notion.so/erler-brain
What lives here
Meeting notes, status updates, customer voice, people pages, deep dives with numbers, daily briefings. Hand-edited, single source of truth for data.
When it goes here
After every meeting, briefing, or decision worth team visibility. Default destination for prose with numbers.
Layer 3 · Versioned Code
GitHub
chris1928a/erler-brain-v3
What lives here
Skills, frameworks, prompts, checklists, reference files, scripts. Version-controlled, CI-checked, single source of truth for code.
When it goes here
When a workflow has happened 3+ times and is worth skill-ifying. Then I codify it as skills/<name>/SKILL.md.
Layer 4 · Originals
Google Drive
Erler Ventures shared
What lives here
PDFs, decks, sheets, contracts, financials, call recordings, long-form content. Read-only originals, never derived data.
When it goes here
After creating or receiving a finished document. Linked from Notion or GitHub as the canonical original.
Real example
Diana, Lucanet call this week
Memory/memory/diana_lucanet.md updated: new quota (1M ARR), boss (Tobi), current frustration (territory plan delayed).
NotionDiana's People Page got 3 fresh action items, decisions, and the date of our next sync.
GitHubNothing. No repeatable skill pattern emerged from this one call.
DriveCall recording archived to /people/diana/calls/2026-05-09.m4a.
The rule: no layer auto-feeds another. /audit-memory + /audit-skill run weekly to catch drift. Memory is the cache. Notion + GitHub + Drive are the truth. Mixing them creates the April 2026 13-month-drift incident I never want to repeat.
EO AI Exchange #1
17 / 51
Act 2 · The Discovery
A few months in, this hit me
The same brain
works for entire organizations.

My personal twin handles 25% of my own work. Then I realized: the same Markdown skills, the same harness, the same deterministic glue, can be deployed for an entire team. Personal Brain to Org Brain. Same architecture, different scale.

EO AI Exchange #1
18 / 51
An Org Brain · how it works

An organization has a brain too. Same four parts.Long-term memory is what the org has ever decided, written, or recorded. Short-term memory is what the team loads for this question. Working memory thinks. The body acts. Right column shows what happens when a founder asks the brain to audit 17 processes.

Layer 1 · Long-term Memory of the Org
Everything the org has ever done.
Like a company's tribal knowledge, except actually written down. Survives a person leaving. Survives a quarter ending.
10 org skills + 17 process docs + 3 years of customer engagements + Slack history + Moco time-tracking + Dropbox originals.
In this run
Founder types: “Audit our 17 processes for the next ops hire.”
Nothing from the 3-year archive loads yet. The brain just knows what was asked.
Layer 2 · Short-term Memory
What the org loads for this question.
Like your team pulling the right files before a meeting. Just enough context for the task at hand.
Different access per role. Founder sees everything. Employee sees a scoped subset.
What loads, exactly
The audit skill, the impact-effort framework, 17 process docs from Notion, last 3 months of Slack summary.
About 20 files. The team's tribal knowledge, pulled into focus.
Layer 3 · Working Memory
The org thinks, right now.
Like a senior associate reasoning through the audit. Without the meetings, without the overhead.
Claude Code Bypass in the cloud. Audit-logged per query. Run by founder, scoped by role.
What happens, right now
Reads the loaded files. Walks all 17 processes. Scores impact vs effort for each.
Identifies the top 3 bottlenecks. Calculates founder-time saved per process per week.
Layer 4 · Body of the Org
Where the org actually acts.
Like the team's hands and voices. Real APIs, real outputs, predictable execution.
Slack, Notion, Dropbox, CRM, internal APIs, the team's actual stack.
What the team gets
An Impact-Effort matrix written to Notion. Top 3 bottlenecks posted to #ops Slack.
Team picks up the rollout the next morning. Founder stays out of the loop.
Same shape, scaled to a team. Personal Brain runs on a Mac at 150 EUR/mo. Org Brain runs in the cloud at under 1k EUR/mo. The brain belongs to the org, not to the operator.
EO AI Exchange #1
19 / 51
An organization remembering

How an org actually remembers.Most orgs lose memory in Slack threads, calls, and email. The standard fix in 2026: two capture modes. Humans update what needs judgment. Cron captures what does not. AI bridges them. Destination is always Notion, GitHub, or Drive, never a single operator's head.

80%
of org memory dies in Slack, email, and meetings. The remaining 20 percent lives in stale Notion pages no one trusts. This pattern catches both.
What is cron?
Cron is a scheduled task that runs automatically on a fixed timetable. Think of it as a recurring calendar event for your computer. “06:00 cron” means: every day at 06:00, this runs by itself. No human pushes a button. No reminder needed. The brain wakes up, does the job, goes back to sleep.
Mode 1 · Humans update
For anything that needs judgment.
Strategy decisions (the why)
People insights (this customer trusts us because...)
Postmortems and retros
ICP and voice-of-customer interpretation
Architectural decisions (ADRs)
Onboarding and role descriptions
Notion (truth)
Mode 2 · AI extracts
Where unstructured becomes structured.
Meeting transcript action items People Page
Customer call voice-of-customer ICP refresh
Slack thread decision log Decisions Page
Email thread relationship update
Recurring pattern new skill candidate
Notion + GitHub (skill candidates)
Mode 3 · Cron captures
For state that should never depend on a human typing.
Code commits (Git hooks)
Deal pipeline state (CRM API pull)
Time tracking (Moco/Toggl)
Support tickets (Zendesk events)
Billing events (Stripe webhook)
Call recordings (Fathom/Granola)
GitHub (code) + Drive (originals) + Notion (state)
Real flow
12-person agency, morning cron
06:00Cron picks up yesterday's call recordings, Slack threads, and Linear updates.
06:05/audit-call + /audit-process extract action items, voice-of-customer, decision signals.
06:10Drafts written to Notion per-customer + per-meeting page. Marked for human review.
08:30Founder reviews top 3 in morning brief, 10 min total. Approves or edits, commits.
Weekly/audit-org-memory flags drift, surfaces skill candidates if a pattern repeats.
The 2026 standard: 80 percent of capture is automated. 20 percent of curation stays human. The 20 is where org judgment lives. The 80 is where org reality lives. Both must be true for the org to actually remember, not just to claim it does.
EO AI Exchange #1
20 / 51
The pattern, side by side

Three bodies. One brain pattern.Long-term memory plus skills on top. Working brain in the middle. Body of tools at the bottom. I run this across one personal life and two organizations. Same shape every time, different deployment surface.

19 Skills Claude Code 39 Workspace tools Telegram + Cron Personal 6 hats, 0 employees 10 Skills Claude Code Bypass GitHub + Notion + Dropbox 17 internal processes Agency 12 employees 8 Skills AWS Lambda Close + Sheets + Slack 5 EventBridge schedules Sales Org 4 reps, 0 managers
Top: Long-term memory + skills
What the brain remembers, plus the procedures it knows how to run. 90 percent of the value. Travels across every deployment.
Middle: Working brain
What thinks, right now. Same role, different surface. Mac, Cloud, Edge, swappable.
Bottom: Body of tools
Where the brain acts. Real APIs, real outputs, predictable execution. No judgment, no vibes.
EO AI Exchange #1
21 / 51
Act 3 · Org Brain 1 of 3 · 12-person Agency
Org Brain · 12-person agency
A 12-person agency under revenue pressure.

Anonymized real case. 17 internal processes. Down 15-20% revenue Q1. Founders considering cuts. Brain inserts before the cut, not after.

Team size
12 employees, 2 founders
Process audited
17 internal processes (sales, ops, fulfillment)
Tools migrated
N8N workflows → Claude Code orchestration
External agency quote
56-89 person-days, 110-180k EUR for V1 build
Brain-supported scope
18-23 PT, 27.5k EUR cap, 1 Junior Dev hire
Source of truth
GitHub agency-brain > Notion > Dropbox (token efficient hierarchy)
Headcount Counterfactual
Without this: 1 Operations Manager (DACH mid-market 80k+) + extended agency engagement. The brain shifts the next hire from mid-senior to junior and keeps founders in the operating seat.
Agency Workspace / Process Audit Q2 / [client redacted]
/audit-process · 17 processes ranked
Low effort
Mid effort
High effort
High impact
Mid impact
Low impact
01Client onboarding handoff, sales to delivery6.2h / wk
02Invoicing follow-up sequence, 30-60-904.5h / wk
03Project status reports to retainer clients3.1h / wk
EO AI Exchange #1
22 / 51
Agency Brain · how it wires

GitHub master. Notion shared. Dropbox raw.Token-efficient hierarchy: code lives in Git, prose lives in Notion, originals in Dropbox.

2 Founders strategic prompts 10 Employees scoped queries agency-brain Claude Code Opus 4.7 Bypass mode, scoped permissions CLAUDE.md per role GitHub agency-brain (master) code + skills + voice rules + audits 10 skills, version-controlled Notion (shared prose) 17 process docs, customer briefings read by team, written by brain Dropbox (raw archive) PDFs, decks, client-confidential read-only, never the source /audit-process Impact-Effort matrix /draft-customer briefing, voice-locked /weekly-report to founders, not to all
EO AI Exchange #1
23 / 51
Act 3 · Org Brain 2 of 3 · B2B Sales Org
Org Brain · Sales Org, Own Build
A B2B sales org, 4 reps, 0 sales managers.

DACH B2B tech recruiting space. Standard playbook would say hire a Sales Manager + a Sales Ops Analyst. We replaced both with AWS Lambda and 5 EventBridge schedules.

Team
4 reps + 1 co-founder, 0 dedicated sales management
Harness
AWS Lambda + EventBridge, 5 schedules Mon-Fri
Bot scoring
Every cold call scored on Gong-style metrics, daily
Coaching delivery
Slack #sales-training, automated drill sheets per rep
Master Dashboard
KPIs auto-built, manual layer preserved (chris-manual)
Co-founder time
From 40h/wk management to 1h/wk review
Headcount Counterfactual
Without this: 1 Sales Manager (80-120k DACH) + 1 Sales Ops Analyst (60-80k) = 140-200k EUR per year. Two hires that never happened, replaced by code that runs at 09:00, 13:00, 18:00, 18:30, 19:30.
Recruitment Circle Workspace
#sales-training 8 members
C
Coaching Bot Today at 09:00
[rep redacted] · daily coaching
Top fail yesterday: jumped to features at 1:43 in the [customer redacted] call. Missed the budget signal at 1:38.
Drill today: pre-pitch discovery, 3-Whys. 15 min before the first call.
EO AI Exchange #1
24 / 51
Sales Org Brain · how it wires

AWS Lambda. 5 schedules. Zero managers.Same 3-layer pattern. Different harness because cron-at-scale lives better in cloud than on a Mac.

Close.com CRM calls, leads, pipeline Gong-style transcripts call recordings, scored Google Sheets Master Dashboard, manual Email + LinkedIn outreach sent + replies AWS Lambda rc-kpi-bot-daily EventBridge schedules 09:00 morning kickoff 13:00 mid-day check 18:00 daily wrap 18:30 lost-reasons 19:30 friday rollup /score-call talk ratio, monologue /daily-coach drill per rep, per pain /lost-reasons pattern detection Slack #sales-training auto-coach posts daily Sheets KPIs auto-built daily manual layer preserved Co-founder Slack weekly summary 1h review, not 40h ops 2 unfilled hires (Sales Manager + Sales Ops Analyst) = 140-200k EUR per year
EO AI Exchange #1
25 / 51
Add it all up

Equivalent-capability payroll 330 to 490k EUR per year, never paid.Same three organizations. Same output. The brain does the job that would otherwise need ~5 FTE worth of headcount.

Organization Equivalent capability EUR / year Actually on payroll
Personal 0.5 EA + 0.5 Analyst + 0.5 Content Mgr 80-110k 0
12-person Agencyanonymized Operations Manager + external agency build 110-180k 0.5 Junior Dev
B2B Sales Organonymized Sales Manager + Sales Ops Analyst 140-200k 0
Total ~5 FTE of capability replaced 330-490k 0.5 FTE
Read the bottom row. About 5 FTE worth of capability, 330 to 490k EUR per year, replaced by the brain. Personal column scaled down to half-time hires because that is what my real ops load justifies, not three full-timers.
EO AI Exchange #1
26 / 51
Single number on screen
Equivalent-capability value, three organizations, one brain pattern
330490k EUR
Per year. Capability I never had to write a job description for, never had to defend in budget review, never had to fire when the cycle turned. Honest equivalent, not a full-time-hires fantasy.

This is what scales. Not the number. The fact that it is the same blueprint. And tonight, on GitHub, the blueprint becomes yours.

EO AI Exchange #1
27 / 51
Tonight, on GitHub
Fork tonight.
Personal Monday.
Org by week 4.
github.com/chris1928a/eo-ai-exchange
Day 1-7
Personal Brain bootstrap
CLAUDE.md, 5 starter skills, memory-curator. Run on your laptop.
Day 8-21
First org skill
/audit-process or /diarize-person on a real org problem. Manual first, then codify.
Day 22-30
Cron live
Move from on-demand to scheduled. The harness wakes up without you.
EO AI Exchange #1
28 / 51
Your first weekend

Five steps. One working brain.Time-boxed for a weekend. Real commands. The shortest path from this slide to your first auto-delivered Monday brief.

30 min
Tonight, right after this call
Clone the repo. Read MASTER_TEMPLATE.md once, top to bottom.
git clone github.com/chris1928a/eo-ai-exchange && cat skills/MASTER_TEMPLATE.md
1-2 h
Saturday morning · write your CLAUDE.md
50 lines max: who you are, your hats, your voice rules, your tools. Pointer-only, no content. This is the router.
cp templates/CLAUDE.md.example ~/.claude/projects/YOUR_BRAIN/CLAUDE.md
1 h
Saturday afternoon · fork your first skill
/morning-brief is the easiest win. Customize the prompt for your inbox + calendar + 3 priorities.
cp skills/universal/morning-brief.md ~/.claude/projects/YOUR_BRAIN/skills/
1-2 h
Sunday · wire up the harness
Telegram bot + Claude Code on your laptop OR a small VM. Set a cron for 7:00 Monday.
follow setup-telegram-bot.md in the repo
Monday
07:00
First brief delivered to your phone
You did it. 5h of weekend, your brain runs the rest of the year. Add one new skill every fortnight from here.
Skip these 3
  • Do not install 5 tools first. One workflow, one skill, one win. The stack grows after the first run, not before.
  • Do not write a 2000-line CLAUDE.md. 50 lines, pointer-only. Long CLAUDE.md kills model attention and your weekend.
  • Do not automate before manual works. Run /morning-brief manually three times. Only then put it on a cron.
Stuck during the weekend?
Drop in the chat tonight or on Slido #3261743 within 24h. Every broken step you report becomes a new solution in the repo.
EO AI Exchange #1
29 / 51
What to take with you

Five lines.

"
One brain, two scales.
Personal Twin saves the operator. Org Brain saves the company.
Same blueprint as your own body: long-term memory holds the skills, working memory thinks, body executes.
The pattern is yours, tonight.
  Markdown + Git + agentic search beats Vector-DB at personal-brain volume.
  Drive is for originals. GitHub for code. Notion for shared prose.
  If you ask your system for the same thing twice, you failed. Skill-ify it.
  Audit is one day. Repair is one or two. Lock-in is forever.
EO AI Exchange #1
30 / 51
Reference material from here on
End of main story
Appendix.
Personal stack deep cuts.

The 20 slides that follow are the receipts: how every layer of my Personal Brain is actually built. Skip if Slido is hot. Open if Q&A goes deep.

EO AI Exchange #1
31 / 51
Recap before the deep dives

Six tools. Four brain layers.Quick recap of slide 11. The next 14 slides break each tool open. Each one carries its body-association badge in the corner.

Layer 1 · Long-term Memory
gitGitHubcode + skills
ntNotionshared prose
gdGoogle Driveoriginals only
Layer 2 · Short-term
mdMemory files125, on demand
Layer 3 · Working
ccClaude Codethin harness
Layer 4 · Body
tgTelegram botmobile interface
mcp39 MCP toolsGmail, Cal, etc.
cr7 cron jobsautopilot

Push intelligence up into Skills. Push execution down into deterministic tools. Keep the harness thin.

EO AI Exchange #1
32 / 51
Layer 1 · Long-term Memory · Tool 01 · GitHub
Code + Skills + Brain
GitHub.
Where my brain lives in version control. Three repos cover everything: skills, memory, slides, build scripts.
"Every skill I write is a permanent upgrade. Build it once, runs forever."
Active Repos
erler-brain-v3
Master repo. Skills, memory mirrors, build scripts, content drafts, slides, innovation patterns. 1.72 GB, 58k files.
private
eo-ai-exchange
This event repo. Open-source solution pipeline. Slides, skills, speakers.
public, MIT
cape-town-ai-exchange
Cape Town hub. Skills, quickstarts, meetup notes for the local builders community.
public, MIT
Path structure
/skills/
17 cluster skills (aviation, real-estate, content, paid-blueprint-engine, parkside-pricing, etc.)
~200 .md files
/memory/
125 diarized 1-page profiles, mirrored from ~/.claude/projects/
git-tracked
/innovation/
Architecture essays, pattern research, group chat insights
live notes

Anti-pattern I hit and removed: OneDrive + Git. Index corruption risk because OneDrive syncs .git/objects mid-write. Repo lives in C:\Code, never in OneDrive.

EO AI Exchange #1
33 / 51
Layer 1 · Long-term Memory · Tool 02 · Notion
Shared Knowledge
Notion.
Where I share with collaborators. Voller Inhalt, never link-rot via .md references. The brain hub for non-developers.
"Skills + frameworks live in GitHub. Meeting notes, status updates, deep dives with numbers live in Notion."
Brain Hub structure
01
Personal & Family. Wedding planning, travel, life ops
private
02
Ventures. Erler Ventures, every active product
private
03
Real Estate. 7 properties, holdings, master file
private
04
Aviation. SACAA training log, NAV logs, exam prep
private
05
Content. Blueprint chapters, weekly engine, podcast
selective sharing
09
Innovation & AI Research. Patterns, group chat insights, vortrag backbones
selective sharing

The rule: Notion holds the full content, never just a "see file X.md". Drive holds originals, GitHub holds code, Notion holds the prose.

EO AI Exchange #1
34 / 51
Layer 1 · Long-term Memory · Tool 03 · Drive
PDFs + Decks
Drive.
Source-of-truth originals only. Not a brain anymore. That was a costly lesson.
"Drive-as-brain failed at API rate limits, OAuth refresh, 200 to 2000ms latency, and constant drift. ripgrep returns in <50ms."
What lives in Drive
/contracts/
Real estate purchase agreements, loan docs, partnership agreements
PDFs
/decks/
Pitch decks, sales materials, board reports
PPTX
/legal/
Insolvency docs, prenup drafts, dispute correspondence
PDFs
/aviation/
SACAA exam prep PDFs, AIPs, airport approach charts
official docs
/financial/
DATEV uploads, bank statements, tax filings
monthly

Migration learning: v4 of my brain tried to put memory + skills in Drive. Six weeks of half-built sync. Killed it April 27. Drive is now read-only originals, brain lives local + GitHub.

EO AI Exchange #1
35 / 51
All 17 Clusters

Skills inventory.

/aviation
SACAA pilot training
Sub-skills: pre-flight (FPL filing), in-flight (NAV logs, radio), exam (PPL Air Law)
/real-estate
Property portfolio ops
Sub-skills: investments (strategic), parkside-pricing (daily Holidu cron)
/content
Substack + LinkedIn engine
Sub-skills: paid-blueprint-engine, weekly-content-engine with anti-AI voice clone
/people
Personal communication
Sub-skills: personal-coach, investment-coach, wedding-assistant, travel-planner
/utils
Cross-cutting helpers
Sub-skills: whatsapp-chat-automation, voice-to-calendar, draft factory
/invoice-creator
Bookkeeping automation
PDF gen, DATEV upload, monthly receipt sweep, missing invoice tracker
/ev-assistant
Founder OS
Positioning, dual ICP, services, outreach, content calendar (most valuable, see slide 11)
/cold-case-podcast
Podcast production
9-phase episode workflow, source verification, Dan-Carlin-style narrative
/mvp-tester
Product validation
8-phase validation, market research, scoring, content for launch (replaces 38h work)
+30 plugins
Anthropic-shipped
docx, pdf, pptx, xlsx document specialists, plus skill-creator, mcp-builder
/audit-skill
Drift detection
Compares GitHub reference files vs live state, reports deviations
/sync-skill
Reference regeneration
Regenerates references from Notion, commits with diff approval
EO AI Exchange #1
36 / 51
The Discipline

How I build a skill.The OpenClaw discipline, applied.

01
Manual first, 3 to 10 items.
Do the work yourself. Not for a week. For months. For years if needed. Identify what dimensions actually matter, not what feels obvious.
Months to years
02
Document the patterns.
Translate into a SKILL.md. Method-call shape with parameters. Inputs, outputs, scoring criteria, edge cases, banned phrases. 1-2 pages, never 20.
Half a day
03
Automate or cron.
Skill runs on demand via Claude Code. Recurring pattern? Put it on a cron. Parkside daily pricing, weekly content, monthly invoices.
A few hours

The test: if I have to ask my system for the same thing twice, I failed. Not the model. Me, for not codifying it the first time.

EO AI Exchange #1
37 / 51
Skill Deep Dive

My most valuable skill.The founder OS that runs my venture: positioning, ICP, services, outreach, content.

Founder OS, daily use
/ev-assistant
Loads my positioning, dual ICP, services + pricing, voice rules. Generates LinkedIn drafts, outreach per ICP segment, sales prep, content calendar, deal-stage-aware follow-ups.
2
ICP segments locked
3
Service packages
~6h
Saved per week on outreach + content
0
Customer data, only my venture context
# SKILL.md, ev-assistant name: ev-assistant type: founder-os load_when: erler ventures, positioning, ICP, outreach, content ## Purpose Founder operating system for my venture. Positioning, ICP, services, sales prep, content drafts. ## Inputs (4 sources) - Notion: Brain Hub > Ventures > Erler Ventures (live) - Memory: ev_positioning.md, ev_icp_dual.md, ev_services.md - Voice rules: writing_rules.md (anti-AI, deutsche Umlaute) - Calendar + Gmail: deal-stage context per prospect ## Outputs - LinkedIn post drafts in my voice (Anti-AI clone) - Outreach emails per ICP segment (PE vs Founder) - Sales conversation prep with prospect context - Content calendar suggestions - Pricing + service page updates ## Non-negotiable rules No AI-fluff (delve, leverage, harness, robust). German Umlaute mandatory. No em-dashes. Voice match exact. No vendor pitch tone.
EO AI Exchange #1
38 / 51
Live Trace

ev-assistant in action.End-to-end outreach for a PE Operating Partner. ~30 seconds.

Trigger
"Draft outreach to Sandra K, Tier-1 PE OperatingPartner, recently announced PortCo CEO appointment in B2B SaaS."
Skill loads
Memory: ev_positioning.md (operator-architect frame)
Memory: ev_icp_pe_segment.md (PE-specific tone, references)
Voice: writing_rules.md (anti-AI, no em-dashes, Sie-Form for PE)
Calendar: any prior touch with Sandra? (none found)
Gmail: any thread? (none found)
Reasoning
Cold prospect. PE OP. New PortCo CEO trigger = operational gap pain. Lead with diagnostic offer (5-day PE Diagnostic), not full engagement. Reference: operational alpha, founder-dependency score, exit-readiness index. Length: 3 sentences max.
Output 1, LinkedIn DM
Hi Sandra, gratuliere zur CEO-Berufung bei [PortCo]. Bei neuen CEO-Setups in B2B SaaS sehe ich oft 6 bis 9 Monate Reibung beim Operational-Hand-Off, was im Exit auf die Multiple drueckt. Falls fuer dich relevant, mein 5-Tage PE Diagnostic findet die 3 groessten Founder-Dependency-Punkte vor dem Closing. Verfuegbar fuer 30 Min naechste Woche?
Output 2, Email follow-up
Auf LinkedIn habe ich gerade kurz angefragt. Falls du den Kontext lieber per Email hast: ich arbeite mit PE OPs an der Bridge zwischen Founder-Dependency und Exit-Readiness. Drei Cases, alle DACH, alle mid-market. Anbei eine 1-Pager Summary von einem.

What is real here:the trigger, the skill load chain, the voice. What is illustrative:the prospect name and PortCo (anonymized for the talk).

EO AI Exchange #1
39 / 51
The Dual ICP

Two segments. One bridge.

Segment A
PE Operating Partners
Geo
DACH mid-market, >1B AUM
Trigger
New PortCo CEO, growth-stage operational gaps
Pain
Founder-dependency, slow operational due-diligence
Tone
Sie-Form, references, operational-alpha frame
Lead with
5-day PE Diagnostic (productized)
Outcome
Founder Dependency Score, Exit-Readiness Index
Segment B
Series-A Founders, 1 to 10M ARR
Geo
DACH primary, EU secondary
Trigger
Hiring spree, key person risk, Series-B prep
Pain
Burning out, no operating system, can't sleep
Tone
Du-Form, founder-to-founder, lived-experience
Lead with
Founder 12-Month engagement
Outcome
Operator-to-Architect transformation

The bridge is the same operating system. What changes is who I sell it to: the PE OP buys it as risk-reduction, the founder buys it as personal liberation.

EO AI Exchange #1
40 / 51
3 Service Packages

From one-off diagnostic to 12-month engagement.

Tier 01
PE Diagnostic
Audience
PE OPs pre-investment
Duration
5 days
Output
Founder Dependency Score, 3 biggest gaps, 90-day fix plan
Pricing
15K EUR fixed
Margin
High, productized
Tier 02
PE Full Engagement
Audience
PE OPs post-investment
Duration
12 weeks
Output
CEO transition, operating system install, exit-readiness index
Pricing
75K to 150K EUR
Margin
Mid, success-fee optional
Tier 03
Founder 12-Month
Audience
Series-A founders
Duration
12 months
Output
Operator-to-Architect transformation, system install, advisory access
Pricing
60K EUR + equity
Margin
Equity-leveraged upside

Same operating system, three buyer-pain frames. The skill knows which one to load based on prospect signal.

EO AI Exchange #1
41 / 51
Anti-AI Voice Clone

Voice rules.The non-negotiables loaded into every output the skill generates.

Banned
AI fluff: delve, leverage, unlock, harness, synergy, robust, seamless, ensure, foster, empower
Em-dashes (—) and en-dashes (–). Use commas or two short sentences.
Enthusiasm filler: "Großartige Frage!", "Absolut!", "Klar!", "in der Tat", "zweifellos"
Mega-paragraphs. Cap at 3 sentences in DM, 5 in email.
Vendor pitch tone. No "transform", "elevate", "next-level", "game-changing".
Repeating user input as opener. No "Great question, here is what I think...".
Required
German Umlaute (ä ö ü ß) always written, never ae oe ue ss as substitute
Direct, terse, natural. Shorter is better. Punctuation like a human.
Match the segment. PE OP gets Sie-Form + references. Founder gets Du-Form + lived-experience.
WhatsApp drafts: zero asterisks, plain text, lowercase first letter ok.
One concrete number per outreach if possible (50 percent valuation comp, 6 months operational gap, etc.)
Always sign with first name only in DMs. Full sig in formal emails.

Built up over 30 days:~60 banned words now, started at 20. The skill audits its own drafts and self-corrects before final output.

EO AI Exchange #1
42 / 51
Layer 2 · Short-term Memory · Tool 04 · Memory Files

125 files. Loaded on demand, never all at once.

Schreibregeln
Voice rules, Anti-AI words, Umlaute mandate, no em-dashes, no asterisks WhatsApp
12
Brain Architecture
Brain architecture, v3.5 audit + migration, source map
5
Real Estate
7 properties, Holdings, MASTERFILE, Parkside Holidu pricing history index
15
Aviation
SACAA training (54 exam memos consolidated), NAV log format, FPL filing
10
Erler Ventures
Positioning, dual ICP, services, pricing, content calendar, weekly engine
8
Personal & Family
Wedding (Cin Cin, vendors), travel, life ops, ELI personality test
8
Content & Sales
Substack drafts, Cold Case podcast, B2B Sales Exchange notes, GitHub repos disambig
10
Diverses
Calendly link, Generalkonsulat Kapstadt, ad-hoc reference notes
7

Each file is a 1-page judgment, distilled from dozens of documents. Loaded by skill on demand, never all at once. The MEMORY.md index file is 60 lines, points at relevant memory by topic.

EO AI Exchange #1
43 / 51
Layer 2 · Short-term Memory · Memory File Anatomy

A memory file is structured Markdown.

--- name: stephan_big_decision_apr26 description: Big Decision Framework for Stephan exit decision, Hybrid Pfad C 8-Wo Sales-Sprint type: reference originSessionId: 8b99bcca-... --- Stephan Förtsch (EO Cape Town buddy) seit 26.04 Vater Christopher will raus, Bewertung 320-535k. ## Why Stephan ueberlegt Pfade. Gefragt nach Empfehlung. ## How to apply - Pfad A (sofort verkaufen): liquide schnell, niedrige Bewertung, Risiko Schwarzwald - Pfad B (warten): unklare Variabel, 12+ Monate - Pfad C (8-Wo Sales-Sprint Hybrid): Liquiditaet + Optionalitaet, klare Stop-Loss-Regel - Empfehlung: C wenn Vater einverstanden ## Anchor - Quarterly Review Stephan vor Q3
Frontmatter: what it is, when to load it, type (reference / project / feedback). Skills can find by type.
Why section: the trigger context. Why I built this memory in the first place.
How to apply: the actionable distillation. This is the diarization. Not the raw conversation, not the meeting notes, the judgment.
Anchor: when to revisit. Auto-pruned if stale.
Live trigger: /coach me through Stephan loads this file, system starts at last context, not at zero.
EO AI Exchange #1
44 / 51
Layer 3 · Working Memory · Tool 05 · Claude Code
Thin Harness
Claude Code.
The harness that runs the model in a loop. Reads files, manages context, enforces safety. ~50 line CLAUDE.md, not 20.000.
"My CLAUDE.md was 20.000 lines. The model's attention degraded. The fix was 200 lines, just pointers." — industry observation
My harness layer
CLAUDE.md
~50 lines. Voice rules, anti-patterns, where to find what. Pointers, not content.
always loaded
ripgrep + glob
Agentic search. No vector-DB. Modern models reason over structured text directly.
<50ms
Sub-Agents
Explore, Plan, general-purpose. Parallel agents for research and multi-step tasks.
on-demand
Hooks
Sparingly. PostToolUse on writes, settings updates. Deterministic enforcement only.
deterministic
MCP servers
15+ adapters. Notion, Drive, Gmail, Calendar, Computer-Use, Chrome, Telegram. All optional, loaded per session.
plug-and-play

What the harness does: Loop, files, context, safety. Four things only. Push intelligence up into Skills, push execution down into MCP and Bash, keep the harness thin.

EO AI Exchange #1
45 / 51
Layer 4 · Body · Tool 06 · Telegram
From Anywhere
Telegram.
My brain on my phone. Voice + text triggers, anywhere. Calls back to Claude API with full brain context loaded.
"I read this on my phone in bed before I move. By the time I am at the desk, I know what today is about."
Setup
Server
AWS Lightsail Frankfurt. Sovereign EU stack, GDPR-aware, 20 EUR/month.
3.71.254.207
Bot
39 tools registered. HMAC webhook verification, prompt injection guards (XML tags), TwiML escaping.
production
Memory access
Skills + Memory loaded via Git pull from GitHub mirror, hourly sync.
git-based
My daily triggers
morning brief
Inbox + Calendar + Drive priorities, 1 page in 30 seconds. 7am every day.
5h saved/wk
prep meeting with X
LinkedIn last 3 + past mail + Drive + opener suggestion. 5 min before any call.
8h saved/wk
draft outreach to Y
ev-assistant skill, ICP-segment-aware, voice-rule-compliant. LinkedIn + Email pair.
~6h saved/wk
log this call
Voice memo to memory file, structured by skill, indexed by topic + person.
diarization

The interface: Telegram is where I talk to my brain when I am not at the desk. Voice in, structured output out.

EO AI Exchange #1
46 / 51
What Runs Without Me

Cron workflows.If I have to start it twice, it goes on a cron.

Daily 06:00
Parkside daily pricing. 6 studios, Holidu GraphQL update, 12+ pattern memory consolidated.Reads bookings, competitor prices, calendar, cluster. Writes optimal price + Notion log.
Real revenue
~1h/day saved
Hourly
Drive memory sync. Git pull from GitHub mirror to local + Telegram bot server.Replaces the v4 Drive-as-brain that failed. Now Git-based, deterministic.
Sync
continuous
Sun 14:00
Weekly content engine. LinkedIn post + Substack teaser drafts for the week.Competitive intel scan, anti-AI voice clone, anti-repetitions check, overlap with prior 30 posts.
Content
~4h/wk saved
Mon 1st of month
Invoice generation + DATEV upload. All client invoices PDF, signed, sent, archived.Automatic numbering, client database integration, post-send archive in Notion.
Bookkeeping
~6h/mo saved
Fri 17:00
Weekly review. Past 7 days summary, open threads, next-week priorities, missed actions.Uses memory diarization across the week, builds 1-page snapshot.
Reflection
2h/wk saved
As-needed
Receipt sweep. Caya + Gmail receipts pulled, JPG to PDF, queued for DATEV.Triggered manually but should be cron in v3.6 next iteration.
Bookkeeping
WIP
EO AI Exchange #1
47 / 51
What I Do Not Do

Four anti-patterns I avoid.

×
No Vector-DB over personal brain
Threshold is ~10K files. I have ~200. Anthropic, Y Combinator, Karpathy all dropped vectors for agentic search. ripgrep returns in <50ms. Anthropic Boris Cherny: "Early Claude Code used RAG + vector DB, agentic search works better."
×
No Drive as brain storage
API rate limits, 200 to 2000ms latency vs <50ms ripgrep, OAuth refresh issues, drift between Drive and local. Drive is for source-of-truth originals (PDFs, decks, contracts), not for a brain. v4 of my brain tried this. Killed it April 27.
×
No OneDrive plus Git
Git index corruption risk. OneDrive syncs .git/objects while Git writes. Race conditions, false API errors. GitHub Desktop warns explicitly. Fix: moved repo to local C:\Code in 43 seconds via Robocopy.
×
No mega-CLAUDE.md with everything
Early CLAUDE.md files were 20.000 lines. Model attention degraded. Cut to ~200 lines of pointers. My CLAUDE.md is ~50 lines. Skills + Memory + Resolvers load on demand. Less in the harness, more in the skill files.

SourcesBoris Cherny (Anthropic) on agentic search · Karpathy LLM Wiki Pattern, April 2026 · my own Brain v3.5 audit, April 27 2026

EO AI Exchange #1
48 / 51
Where I Land

Solo founders converge.

Chris Industry pattern Karpathy Anthropic
StorageMarkdown + GitMarkdown + GitMarkdown + ObsidianMarkdown
HarnessClaude Codegbrain CLI ~200 linesCustom PythonClaude Code
Skills formatMarkdown SKILL.mdMarkdown method-callsWiki + backlinksMarkdown
Vector-DBNoNoNoNo
Searchripgrep / agenticgrep / agenticgrep / agenticripgrep / agentic
Skills volume17 cluster + 30 pluginsnot publicminimalistmarketplace
Memory volume~125 filesnot publicunknownn/a
MCP layerYes, 31 toolsnot publicminimalYes
Cron / scheduledYes (Parkside, weekly)not publicnot publicNo
Mobile interfaceTelegram botNoNoClaude.app
Multi-userNoNoNoYes (teams)

Solo-founder setups converge on Markdown + Skills + agentic search. Vector-DB is for 10K+ unstructured files. If your volume fits in your head, your brain fits in Markdown.

SourcesIndustry brain repos · Karpathy LLM Wiki gist · Boris Cherny (Anthropic), "Claude Code uses agentic search, not vectors", April 2025

EO AI Exchange #1
49 / 51
Direct Comparison

Where I align with the pattern. Where I differ.

Industry-aligned
Markdown + Git as source of truth, no Vector-DB
Skills as method-calls with parameters, judgment in Markdown
~50 line CLAUDE.md as resolver, never 20.000 lines
Agentic search via grep, modern models reason over text directly
Diarization as the knowledge-work hack, 1-page profiles
OpenClaw discipline: manual first, document, then automate
Where I differ
Heavy MCP layer: 31 tools across Notion, Drive, Gmail, Calendar, Telegram, Computer-Use. Tan does not publish his.
17 thematic clusters (aviation, real-estate, content, people) plus 30 plugins. Tan focuses on YC ops only.
Telegram mobile interface via AWS Lightsail Frankfurt. Voice + text triggers from anywhere.
Cron-driven workflows: Parkside daily, weekly content, monthly invoices, weekly review.
Notion as sharing layer for non-developers. Tan has not addressed the share-with-others problem publicly.
EU-only stack: AWS Frankfurt, deutsche Umlaute as voice rule, GDPR-aware by default.

Net: The industry pattern is the foundation. My differences come from being a multi-domain operator (real estate + content + ventures + aviation) who lives outside of Mountain View. The architecture is identical, the surface area is wider.

EO AI Exchange #1
50 / 51
Recent Audit + Migration

From chaos to compound, in 30 days.

Mar 18
Brain v4 upgrade Phases 1-5. Drive-Hub as brain vision. Telegram bot security fixes. 39 tools registered. PDF/DOCX parsing. Health checks.
Mar to Apr
30 days of build-up. Skills cluster expansion. ~17 clusters reached. 175 uncommitted files. Drive sync half-finished. Memory bloat from 54 SACAA exam memos still active after passing.
Apr 27 AM
Brain v3.5 Audit. Drive sync killed. OneDrive + Git race condition flagged. ~10K vector-DB threshold confirmed not met. Vector-DB definitively skipped.
Apr 27 PM
Migration in 1 day. Robocopy to C:\Code (1.72 GB / 58.638 files in 43 sec). Memory-folder copied to ~/.claude/projects/. 11 legacy skills archived. MEMORY.md compressed 124 to 60 lines.
Apr 27 EOD
Architecture documented. Vortrag backbone written for this talk. Cross-references to Tan, Karpathy, Anthropic. Roadmap May to September drafted.
May 11
This talk. Architecture as artifact. Anti-patterns shared. EO operators get the playbook before they hit the same walls.

Audit is one day. Migration is one day. Documentation is one day. The cost of architecture drift is always shorter to fix than to live with.

EO AI Exchange #1
51 / 51