← → navigate · space next
Talking to Machines

Day Three.

The day it becomes yours.

01 · Welcome
Day 2 →

The map of AI — in one glance.

The Big Three

ChatGPT · Claude · Gemini

Claude is the current choice. Gemini is worth a try.

Open weights · China

Qwen · DeepSeek · Kimi

Frontier quality at a fraction of the cost.

On your own laptop

Gemma · LM Studio · Ollama

Private. Free forever. Offline. No rate limits.

The rest of the map

Must-use · Try · Avoid

  • Must-use NotebookLM · Nano Banana · Lovable
  • Try Runway · ElevenLabs · Suno / Udio
  • Avoid AI browsers
02 · Day 2 recap · the map
Day 2 →

An agent is just an LLM in a loop with tools.

01 · Read
Think
→
02 · Pick a tool
Act
→
03 · Read result
Observe
↻
04 · Until done
Repeat
How they reach the world
Terminal
CLIs
Universal plug
MCPs
Any service
APIs
03 · Day 2 recap · agents
Day 2 →

Your first piece of personal software.

We built Chrome extensions.

The first attempt had errors. We iterated.

And we experienced, first-hand — you don't need to be a developer to make software.

04 · Day 2 recap · extensions

Let's build on Day 1 + Day 2 and make a 90-day plan.

Beat 01

Skills, in depth.

Your way of working — captured, given to AI. Demo the best ones out there. Build one together.

Beat 02

The 90-day plan.

A real plan for your business, your life, your goals. Know · Do · Show — laddered across 90 days.

05 · Today
First — before the plan →

Skills.

Train · program · augment AI to work your way

06 · Break
What is a Skill, really

Your way of working — captured in a folder.

≈ Skills are memory, distilled. Every lesson worth repeating, pressed into a file any agent can carry.
Part 01

The brief

One line that tells Claude when to step in. "Use me whenever I'm writing an investment memo."

Part 02

The playbook

How you actually do the work. Your steps, your tone, examples of what good looks like — and what doesn't.

Part 03

The toolkit

Optional. Templates and shortcuts Claude can reach for — your deck template, your Excel formulas, your boilerplate.

Write it once. Every future session inherits it.
07 · Skills anatomy
Where they live

Finding good skills.

Source 01 · Anthropic

The official library.

Anthropic publishes production-grade skills on GitHub. Financial services, Office suite, brand guidelines. Safe to trust.

github.com/anthropics/skills

finance pptx xlsx docx pdf
Source 02 · Community

Curated by others.

Developers share skills on GitHub, X, and Discord. Quality varies wildly. Read every file before installing — a skill can run code on your machine.

search GitHub "claude skills"
Source 03 · Yours

The ones that matter most.

The skill for your investment memo. Your tone of voice. Your weekly report. No one else can write these. They're your moat.

we build one · next
Workshop pack A head start — 13 skills to try. Research, decisions, audio, browsing, prompts, more. Download ↓
08 · Finding skills
Build one with me →

Ask Claude to make the skill for you.

The Skill Creator is Anthropic's own skill for writing skills. It asks you the right questions, writes the files, validates the trigger, and installs it. All while you talk in plain English.

The prompt → 
Use the skill-creator skill to help me build a Skill.

I send an investor update on the last Friday of every month.
Same structure every time: cover · highlights · metrics ·
one surprise · one ask.

Interview me about my voice, format, and what "good" looks like.
Then draft the SKILL.md, a good example, and a template.
09 · Skill creator
Now — the main event →

The 90-day plan.

the ultimate takeaway · a plan for Monday

10 · Break
01
Phase 01 · Weeks 1 → 4

Stop one-shotting.

The Loop, into muscle memory →
Week 01

Scan the field.

  • Run the same task through Claude, ChatGPT, and Gemini. Where does each one win?
  • Tag every real task as Know · Do · Show before you start.
Week 02

Prep before you prompt.

  • Ask AI to interview you before it starts the task.
  • Ask AI to rewrite your prompt before you run it.
  • Keepcross-check 2+ models
Week 03

Direct, don't dictate.

  • Show good and bad examples from your past work.
  • 3–5 passes, minimum. No first answers.
  • KeepAI interviews you first
Week 04

Don't ship raw.

  • Push back. Ask what AI held back. Ask what would make it better.
  • Humanize. Deslop every edge.
  • Keepexamples · iteration · cross-check
11 · Phase 1
02
Phase 02 · Weeks 5 → 8

Sessions become your SOP.

Your corpus becomes Skills →
Week 05

See your patterns.

  • Review your best Month 1 sessions. Where did you actually get leverage?
  • Ask AI to help you spot the patterns.
  • Pick your top 3 recurring tasks — one Know, one Do, one Show.
Week 06

Write one SOP.

  • Write your most-repeated task like you'd teach a smart intern.
  • Have AI interview you to surface edge cases you'd forget.
  • Add examples of good and bad output.
Week 07

Ship one Skill.

  • Skill Creator turns your SOP into a Skill.
  • 3–5 passes until it works on 3 real inputs.
  • Exploreone AI tool you've never used, on a real task
Week 08

Compose.

  • Build a second Skill in a different K/D/S category.
  • Chain two Skills in one real workflow.
  • Hand a Skill to a colleague. Watch what breaks.
12 · Phase 2
03
Phase 03 · Weeks 9 → 13

Skills are your operating system.

The foundation everything runs on →
Week 09

Hunt for Skills.

  • Browse anthropic/skills, skills.sh, community lists.
  • Install two Skills you didn't write. Use each on a real task.
  • Ask AI: "What Skills should I be using?" Cross-check with a second model.
Week 10

Scale your library.

  • Add two more Skills in the K/D/S categories you haven't covered.
  • AI interviews you before the SOP.
  • ExploreNotebookLM · Nano Banana · Deep Research · Perplexity · local model
Week 11

Compose across domains.

  • Chain 3+ Skills end-to-end (Research → Draft → Polish).
  • Watch where handoffs break. Tighten inputs and outputs.
  • Exploreone "must-use" from the Day 2 map you haven't tried
Week 12

The meta move.

  • Build a Skill for building Skills — your taste, baked in.
  • Run it on your existing Skills. What gets sharper?
  • Exploreswap your default chat AI for a week
Week 13

Share, audit, teach.

  • Publish one Skill. Install one more from someone else.
  • Stack audit: keep / drop / add — Skills, tools, sources, habits.
  • One-pager: your Skill library, one line each.
13 · Phase 3
Case study · the receipts

Ramp — the company that went all-in.

$22B fintech · ~1,200 people · 4 quarters of accelerating revenue →
99.5%

of ~1,200 Ramp employees are active weekly AI users. Not just engineers — finance, ops, marketing, recruiting. The whole company.

"If you're not using Claude Code this year — no matter what your role is — you're probably underperforming compared to others in the company. The people still in L0 will most likely not be at the company." — Geoff Charles, CPO · April 2026
What's actually running inside
  • ~30% of merged PRs written by Inspect — their internal background coding agent. Sandboxed VMs, pre-built images, runs in Slack, VS Code, Chrome, PRs, voice.
  • 1M+ lines of AI-suggested code shipped in 30 days. 50% weekly Claude Code adoption across engineering.
  • 80% reduction in incident investigation time — Claude Code plugged into Datadog, Sentry, logs, flags.
  • 84% of non-engineers use coding agents weekly. Finance, sales, risk, recruiting — all query Snowflake in plain English.
  • 350+ AI workflows shared peer-to-peer on their internal Skills marketplace. Plus a 3-person team (Jolt AI) acquired to build the platform.
The ladder Ramp is measuring everyone on →
L0
Occasional chat.

"I tried ChatGPT once." Nothing sticks.

L1
Daily basics.

AI a few times a week. Drafts, quick answers.

L2
Embedded in workflow.

Every role has 3+ AI-assisted tasks. Daily.

L3
Systems do the work.

Agents run real work. Humans direct.

14 · Ramp case study
And here's the beautiful part →

Our 90-day plan = Ramp's L0 → L3 ladder.

Same framework. Same destination. You just don't need a $22B company to start.

Our plan · Phase 01

Stop one-shotting.

The Loop into muscle memory. Cross-check, interview, iterate, humanize.

=
Ramp · L0 → L1

Occasional chat → basic daily use.

From "I tried ChatGPT once" to AI in your workflow. The floor Ramp no longer tolerates.

Our plan · Phase 02

Sessions become SOP.

Spot patterns. Write one SOP. Ship one Skill. Compose two. Hand one to a colleague.

=
Ramp · L1 → L2

AI embedded in daily workflow.

The inflection point. Every role has their 3 tasks AI-assisted. Ramp's "internal marketplace" starts here.

Our plan · Phase 03

Skills are your operating system.

Install, scale, compose across domains. Meta-skill. Publish, audit, teach.

=
Ramp · L2 → L3

Systems that do the work.

At Ramp: hundreds of agents ship 30% of code, handle incidents, build dashboards. At your desk: composed Skills doing your weekly work, overnight. Same shape, different scale.

15 · Ramp × Plan
Your turn — before you leave →

Your plan. Personalized. In 8 minutes.

A prompt interviews you. Claude writes your plan. You walk out with it.

Go here on your laptop
talkingtomachines.xyz/workshop-day3/plan

One "Copy prompt" button. One click to open Claude. 10 quick questions — mostly multiple choice. You get back a printable HTML page, named for your work.

  1. Hit Copy prompt. Hit Open Claude.
  2. Paste into a new Claude chat. Send.
  3. Answer the questions — pick a letter, or type your own.
  4. Confirm the 4-line summary. Claude writes your plan.
  5. Save the HTML. Pin it somewhere you'll see it Monday.
Do it before you leave the room. Tuesday morning is too late.
16 · Take-home