DAY 3 · 70 MINUTES · SHOWCASE ONLY

Edge Technology
AI Showcase

Three days ago your agent didn't exist. Today we see how far it can go — decks by your rulebook, a second brain, a skill store, and skills that refine themselves.

4 showcasesAll liveLaptops can stay closed
Personal Agentonline
09:41
Build the client deck from my notes — follow our Design.md.
On it — card layout, charcoal palette, max 6 cards per slide. Your .pptx is coming. 🎴
second brain · connected 3 related memories about this client
Also: found a ClawHub skill that auto-charts your numbers. Installed, vetted, and I already improved it after a test run.

← → navigate  ·  F fullscreen  ·  ⌘P export PDF

AGENDA

Four showcases, one leap forward

The last three days were about how to build. This session is about how far this goes — four live proofs of what your agent can already do.

SHOWCASE 1 Slide Enhancement
SHOWCASE 2 Second Brain
SHOWCASE 3 ClawHub
SHOWCASE 4 Skills Refinement

It designs decks by your rulebook

  • Design.md → .pptx, live
  • One card, one point

It remembers — and connects

  • Karpathy's second brain
  • Graph memory + vault

It learns from a skill store

  • 362 ready-made skills, live tour
  • Vet → install → use in minutes

Its skills improve themselves

  • Self-improving agents
  • Loops & graphs
Nothing to type today. Watch, note down ideas — every one of these runs on the agent you already built.
HOW TO READ THIS SESSION

The star employee, three days in

We hired a digital employee three days ago. Today: four proofs of how good it got.

D1

The employee was born

Terminal, OpenClaw, a VPS home, guardrails — your agent came alive and answered on Telegram.

D2

It got access to your work

Drive, Docs, Slides, Gmail — the access card to the filing room, plus the written rules that protect it.

D3

This morning: you made it smarter

The Skill Maker Competition — you built a working skill with your own hands. Most professionals never have.

NOW

How far can it actually go?

Decks by the rulebook, a growing memory, new skills from a store, and skills that improve themselves.

Every showcase is the edge of what's possible right now — not a demo video, not next year. Today.
1
SHOWCASE 1 OF 4 · 12 MINUTES

It designs decks by your rulebook

Stop designing slide by slide. Write the design rules once — one Markdown file — and the agent turns any content into a deck that follows them. Output: a real PowerPoint file.

Live demoDesign.md → .pptx
Analogy: a brand book the agent can actually execute — rules on paper become slides on screen
SHOWCASE 1 · THE RULEBOOK, ON SCREEN

One file holds the whole design system

Design.md visual preview
THE DESIGN.MD, RENDERED — theme, components, palette: taste written down as a spec
Typography spec
THE TYPE SCALE — Cambria titles, Calibri body: fonts every office laptop has
  • "One card, one point." No floating text — every idea lives in a card.
  • Max 6 cards per slide. Too much content? Fewer cards — never smaller fonts.
  • A QA checklist is inside the spec — overflow, gutters, margins: the agent checks its own work.
This is the shift: you don't design slides anymore — you design the rules.
SHOWCASE 1 · INSIDE THE FILE

Inside the Design.md — the parts that matter

deck-card-presentation-DESIGN.md
## 1 · CORE CONCEPT
"The card is the unit of thought — one card, one point."
## 4 · LAYOUT GRID
1 card → hero  ·  3 → columns  ·  6 → 3×2 grid
Never exceed 6 cards — if content runs long, reduce cards, not fonts.
## 5 · TYPOGRAPHY
Slide Title — Cambria Bold · 36–40pt
Card Title — Calibri Bold · 18–20pt
Card Content — Calibri · 13–14pt  ·  always left-aligned
## 5 · COLOR PALETTE — pick one theme, keep it
#36454F #F2F2F2 #212121 CHARCOAL MINIMAL TEAL TRUST
accent color is only for icon circles, hero numbers & tags — never body text
## 7 · QA CHECKLIST
☐ no overflow   ☐ equal card heights   ☐ max 6 cards   ☐ uniform radius & shadows

The thought unit

If it doesn't fit one card, it's two ideas. Structure you can see.

The hard cap

Six cards, never more — the rule that protects your audience from slide soup.

The type scale

Sizes are fixed, fonts are office-safe (Cambria + Calibri). Nobody "picks fonts" ever again.

The self-check

The agent verifies its deck against this exact list before handing you the file.

SOURCE Full file: deck-card-presentation-DESIGN.md — in the section's source folder

SHOWCASE 1 · LIVE DEMO

Design.md + messy notes → finished .pptx

1 · Show the rulebook (Design.md)› 2 · Hand over raw content› 3 · Agent builds the .pptx› 4 · Open it in PowerPoint
Personal Agentonline
10:12
Here are my messy notes for the client update. Build the deck — follow our Design.md.
Reading the rulebook: card-based, max 6 cards per slide, Charcoal Minimal palette, Cambria & Calibri. Building… 🎴
client-update.pptx · 12 slides · running the QA checklist from the spec
Done ✓ One section had 7 points, so I split it across two slides — the rulebook forbids more than 6 cards. File attached.
WHY .PPTX?

Because that's what your company actually uses. The deck opens, edits and shares like any other PowerPoint — nobody needs a new tool.

Watch the last message: the agent didn't just obey the rules — it enforced them on your content. That's a design system working.
2
SHOWCASE 2 OF 4 · 20 MINUTES

It remembers — and connects

A good assistant takes notes. An extraordinary one connects them: "you said client X hates morning meetings — so I moved tomorrow's slot."

Second brainGraph memoryKnowledge vault
Analogy: the assistant who never forgets
SHOWCASE 2 · A GUESSING GAME

Before we explain anything — what is this?

mystery graph
Shout it out — what are we looking at?
A galaxy? 🌌 Neurons? 🧠 A social network? 👥 A virus? 🦠

All good guesses. Two hints: nobody drew it — it grew. And it belongs to one person.

SHOWCASE 2 · THE REVEAL

This is one person's second brain

the whole second brain
EVERYTHING ONE PERSON'S AGENT REMEMBERS — seen from above
  • Every dot = one thing it remembers — a bill, an account, a goal, a person, a habit
  • Colors = kinds of things · clusters = areas of life — money, work, home, family
  • The lines = how they relate — and that's where the magic lives
  • Nobody filed anything. It grew from normal chatting, day after day
The idea has a name — a second brain (Tiago Forte made it famous). The classic version dies because you do the filing. Here, the agent is the librarian.
SHOWCASE 2 · ZOOM 1 — THE CONNECTIONS

Zoom in: every line means something

monthly-budget hub connections
THE MONTHLY-BUDGET HUB — every line is a bill, account or habit attached to it
  • Read the labels: listrik-pln, netflix, spotify, cicilan-motor, langganan-gym, zakat-bulanan, bca-mobile. Not tech jargon — life.
  • Hubs emerge on their own: monthly-budget, bca-mobile — the places everything else attaches to.
  • Ask "what subscriptions can I cut, and how much do I save?" — the agent walks these lines to answer.
Memory that connects beats memory that stores. A pile of notes can't answer that question. A graph can.
SHOWCASE 2 · ZOOM 2 — OPEN A NODE

Click a dot: every node is a page

monthly-budget node opened
THE MONTHLY-BUDGET NODE, OPENED — a real page with real structure
  • Properties, not loose notes: type, title, tags, status — every memory is structured.
  • The orange line: generated by claude-code · 2026-08-21 — nobody typed this page.
  • "What feeds into it" — live links to other nodes: Listrik PLN, Air PDAM, Internet WiFi, BPJS…
  • Bottom corner: 65 backlinks — sixty-five other pages point here.
The agent wrote it. The agent maintains it. You just... lived.
SHOWCASE 2 · ZOOM 3 — INSIDE THE NODE

Inside: real numbers, and under it — a file

monthly budget content
THE CONTENT — income Rp40jt · bills 43% · savings 16% · the agent's own analysis, linked to the two goals that rate funds
raw markdown leaf node
UNDER THE UI — a leaf node is just a text file: type Bill · Rp350.000/month · due day 15 · auto-debit via BCA
This is the knowledge vault: claims + evidence + timestamps the agent writes, updates and re-reads. Structure, not vibes.
SHOWCASE 2 · INTERLUDE · 2 MIN 15

Open loops, closed while you live your life

Sechel Agents film
SECHEL AGENTS — a film by Johannes Christian · health · bills · promises · people
THE HEADLINE

100 nodes. 400,000 words. He wrote none of it.

100
nodes in his wiki

on a single research topic

400k
words inside them

written, linked & indexed by the LLM

0
written by him

he only feeds it new material

"You rarely ever write or edit the wiki manually — it's the domain of the LLM. … I think there is room here for an incredible new product."

— Andrej Karpathy (OpenAI founding member, ex-Director of AI at Tesla), on his self-writing research wiki

You're not waiting for that product. The agent you built this week already works exactly like this.
SHOWCASE 2 · WHY "GRAPH", NOT JUST "NOTES"

Three levels of memory

LEVEL 1

Chatbot — remembers one session

Close the tab, it's gone. Every conversation starts from zero.

LEVEL 2

File memory — saved notes

It remembers... like a pile of sticky notes. Facts survive, but nothing is connected.

LEVEL 3

Graph memory — a network of entities & relationships

People, projects, preferences as nodes; the relationships as edges. It remembers and understands the connections.

Google's Knowledge Graph · 1.6T facts on 54B entities — the same trick, planet-scale Zep research · graph memory: +18.5% accuracy, −90% latency vs raw chat history

SOURCES Wikipedia — Knowledge Graph (Google), May 2024 · Zep, arXiv 2501.13956

Your agent runs Level 3 — the universe you just saw is a few days of normal use.
3
SHOWCASE 3 OF 4 · 15 MINUTES

It learns new skills from a store

A phone without apps can only call and text. What makes it useful is the app store. ClawHub — clawhub.ai, "Claws for your Claws" — is the app store for your agent.

Live democlawhub.aiAll through chat — no terminal
Analogy: an app store for agents
SHOWCASE 3 · THE OPENER

Until now, every skill came from you

1 · By hand — this morning's competition› 2 · By agent — you brief, it drafts› 3 · …but the world got there first

Built by hand

This morning you wrote a SKILL.md yourself, tested it, fixed it. Now you know what a skill costs to build — and that it's the same format the whole ecosystem uses.

Built by your agent

You brief, the agent drafts the skill for you. Faster — but still built from zero, and only for you.

Meanwhile: 362 ready-made skills are sitting on the marketplace — built, tested and shared by people who had your problem first. Let's take the tour.
SHOWCASE 3 · THE STORE, FROM THE LIVE SITE

"Claws for your Claws" — what's on the shelves

clawhub.ai homepage
CLAWHUB.AI — skills & plugins from top creators: trending, featured, official
Skills Hub on myclaw.ai
SKILLS HUB — "extend your OpenClaw with community-built skills" · 362 skills, 15 categories
Gog · 175.5K downloads — wires Gmail, Drive, Docs Humanizer · 98.6K — fixes robotic AI writing ontology · 128K — graph memory (Showcase 2!) Polymarket · 147K
Installing is one line — npx clawhub@latest install humanizer — and your agent runs it from chat.
SHOWCASE 3 · READ THE SHELF LABEL

A skill page reads like a hiring CV

ClawPDF Master skill page
CLAWPDF MASTER — PDF extraction, OCR, merge/split, PDF→Markdown, by @northcap-group

Security audit: Pass

An automated check on every version — helpful, not a blank cheque.

Freshness: updated 13h ago

v1.0.11, live download graph — an actively maintained skill, not abandonware.

Full transparency

SKILL.md, files, diffs, versions, requirements — you can read everything it will do before installing.

Stars, downloads, audit, freshness, source — read the CV before you hire. Exactly like a new employee.
SHOWCASE 3 · LIVE DEMO

Search → vet → install → use. In minutes.

1 · Search by need› 2 · Read the card: author, audit, freshness› 3 · Vet, then install› 4 · Use it in the very next message
Personal Agentonline
14:05
Find me a skill for PDFs — extract tables, OCR, merge reports.
Top result: ClawPDF Master by @northcap-group — extraction, OCR, PDF→Markdown, batch folders. Security audit: Pass · updated 13h ago · v1.0.11.
Vet it first, then install.
verifying source, maintainer, package contents… ✓ · openclaw skills install @northcap-group/clawpdf-master ✓
Installed. Send me any PDF and I'll pull the tables out. Two minutes ago I couldn't do this.
ClawHub vet-first install prompt
THE STORE'S OWN INSTALL PROMPT — "verify its source, maintainer, and package contents… ask before broader changes." Vetting is the documented default.
From "the agent can't" to "the agent can" — in minutes.
SHOWCASE 3 · BEFORE WE MOVE ON

One sentence to hold on to

Someone else's skill = someone else's code, running on your agent, with your agent's access.

The upside you just saw

Capabilities on demand — your agent grows through skills, not by swapping models.

The thing to watch

What if the skill is malicious? Always read what a skill does before you install it.

4
SHOWCASE 4 OF 4 · 12 MINUTES

Its skills improve themselves

Most employees follow the manual. The rare, valuable ones update the manual after every project. Your agent can be that employee.

Self-improving agentsLoop engineeringGraph engineering
Analogy: the employee who rewrites their own SOP
SHOWCASE 4 · FIRST — CONGRATULATIONS

That was version 1. Now watch how far it can grow

This morning you did something most professionals haven't: you built a working AI skill. "Works" is the starting line — from here it only gets better.

Who tested their skill this morning, saw a weak result, and fixed it?
That instinct — test, judge, fix, retest — is exactly what can be handed to the agent itself.
SHOWCASE 4 · THE LADDER

You're already on level 1. Here's the ladder

LEVEL 3

Graph engineering · the frontier

Wire several agents into a team — one drafts, one critiques, one fact-checks, a human signs off. You draw the org chart.

LEVEL 2

Loop engineering · 2026's new meta

Define what "done" looks like; the agent drafts → checks → fixes → repeats until it hits it. You set the target; the athlete does the reps.

LEVEL 1

Prompt engineering · you, since Day 1

Write one clear instruction, review the answer. Briefing a worker, checking every task.

Each level up, your leverage moves: from doing the work → to defining "done" → to designing the team.
LEVEL 2 · LOOP ENGINEERING

Stop being the person who presses "try again"

A chef doesn't serve the first pour — they taste and re-season until it's right. Loop engineering hands your agent the tasting spoon.

DRAFT the agent attempts it CHECK AGAINST "DONE" your definition, not its opinion DELIVER ✓ only when it passes passes fails FIX WHAT FAILED then run it again repeat until it passes the agent goes round this loop by itself — you never press "try again"
"Done" must be checkable — "all numbers verified", "under one page", "no broken links" "The verifier is the bottleneck, not the model"
Your skill moves up a level: from doing the work to defining what finished means.
LEVEL 3 · GRAPH ENGINEERING · THE DIAMOND

When one loop isn't enough, you wire a team

FAN OUT REDUCE SYNTHESIZE SPLIT THE JOB one lead plans the angles WORKER 1 WORKER 2 WORKER 3 WORKER 4 WORKER 5 CHECKER — FRESH EYES never the agent that did the work MERGE INTO ONE ANSWER
Workers run side by side — a graph buys breadth, not better judgment The checker never shares the worker's context
Showcase 2 wired your agent's memories into a graph. This wires workers into one — where "using AI" turns into managing an AI team.
SHOWCASE 4 · MICRO-DEMO

The skill grades itself

One prompt, on a real skill. Thirty seconds of silence while it works is fine — a skill editing itself is the showcase.

agent · skill refinement loop
›
A skill is a living document. Version 1 is the seed — refinement is where the value compounds.
✓
DAY 3 · CLOSING

The agent is alive. Its memory will grow.
Its skills can multiply — and refine themselves.
Where it goes next is up to you.

Everything showcased today runs on your own agent after you go home: the deck rulebook, the second brain, the skill store, the refinement loops.

The tools: from usWhat you build: from you
Thank you for three days of building. Any questions?
1 / 29