Agent Skills Hackathon
Build Your First AI Skill in 40 Minutes
Why You're Here
You'll leave tonight with a working AI skill. Not a tutorial. Something you can use tomorrow.
- You already know what tasks eat your time — now automate them
- A skill is a folder with a SKILL.md file — no coding required
- Zero to working skill in under 40 minutes
- Build alongside people who do completely different jobs
Who's in the Room
- Nurses & clinicians
- Wealth advisors & analysts
- Product managers
- Engineers & founders
- Recruiters & operations
- Teachers & creatives
Different jobs, same problem: repetitive work that shouldn't need a human.
Tonight's Agenda
40 minutes of build time. That's the whole point.
What is an AI Skill?
A folder containing a SKILL.md that teaches an AI agent how to perform a specific task.
Think of it like…
The AI is a chef. It already knows how to cook.
The skill is the recipe. It tells the chef what to make.
Live Demo
Watch a skill get built and used in real time.
↓ Switching to live demo ↓
How We Built That
Three steps from real work to reusable skill:
Voice notes are the fastest path. Record 2 minutes of rambling about what you do — that's enough for a first-draft skill.
Capabilities vs. Preferences
✓ Automatable (Capability)
"Format my meeting notes into a structured summary with action items, decisions, and attendees."
Mechanical, repeatable, structured.
✗ Not automatable (Preference)
"Write in my exact voice with my specific tone and humor."
Subjective — you'll always need to edit.
The Litmus Test: "Do I use the output directly?" If you'd take the skill's output and send it, file it, or present it with only minor edits — it's a good skill candidate.
Skill Folder Anatomy
It's just a folder with a markdown file inside.
my-skill/
├── SKILL.md ← required (this is everything)
├── scripts/ ← optional helper scripts
├── references/ ← optional docs & examples
└── assets/ ← optional images, templatesNo code. No build step. Just a markdown file in a folder.
Inside SKILL.md
---
name: client-follow-up
description: >
Follow-up email after meetings.
Triggers on "follow up,"
"recap," "meeting notes."
---↑ description = how the agent discovers your skill
What the user provides: notes, files, URLs, context
Ordered instructions: 1. Read → 2. Draft → 3. Review → 4. Format
Exact template: show what "done" looks like
Constraints: word limits, tone, edge cases
Key: Include the exact words a user would say in the description. The agent only sees the description at first; full instructions load on demand.
The 5 Skill Killers
- Teaching instead of doing — "Explain what a report is" vs "Write my weekly report"
- Scope too broad — One skill, one job. Split big tasks into chained skills
- No examples — Show what good output looks like, not just describe it
- No output format — Specify structure: "3 bullets, past tense, under 50 words"
- Testing with fake inputs — Always test with your real data
When Things Go Wrong
| Symptom | Likely Cause | Fix |
|---|---|---|
| Agent ignores the skill | Trigger too vague | Be specific in your trigger |
| Output is wrong format | No format specified | Define exact output format |
| Agent hallucinates | Missing context | Add examples of good output to the skill |
| Too slow | Task too broad | Split into multiple focused skills |
| Browser actions fail | Web automation is unreliable | Use APIs instead of web automation |
What Skills Can't Do
- Run standalone — Needs an agent runtime (Claude Desktop, OpenClaw, etc.)
- Be perfectly consistent — Non-deterministic by nature, like a real employee
- Replace human judgment — Great first draft, still needs your review
- Access everything day one — Start small, grant access progressively
Progressive trust: Calendar read → email read → email draft → email send. Grant access incrementally.
Your Skill Suggestions
Based on your backgrounds · Full details in participant guide
Your Skill Suggestions (cont.)
Your Skill Suggestions (cont.)
Pick Your Project
- Weekly status report
- Meeting notes formatter
- Email triage helper
- Client proposal drafter
- Research summarizer
- Expense tracker
- Multi-step sales outreach
- Lesson plan generator
- Clinical note assistant
Build Time
- Stuck? Record a voice note describing your task
- Use APIs before browsers — more reliable
- Test with real data, not examples
- Ask ChatGPT if you need help
Lightning Demos
Celebrate the failures — that's where learning happens.
Beyond Tonight
- Skill chaining — Output of one skill feeds the next (research → draft → review → send)
- Sub-agents — Multiple specialized agents, each in its own lane of work
- Progressive trust — Calendar read → email read → email draft → email send
- Manager mindset — Onboard like an employee: own email, own calendar, scoped role
Key insight: The real reason for multiple agents isn't capability — it's context management. Separate agents stay focused, like Slack channels for your workflows. One agent doing everything = context overload.
Resources
- Skills Master Class (AI Daily Brief)
- Complete Guide to Building Skills (PDF)
- Agent Skills with Anthropic (DeepLearning.AI)
Share Your Skill
Build it tonight. Share it tomorrow.
- Fork the Nerds Like Me repo
- Add your skill as a new folder
- Open a pull request
- Your skill helps someone else's "warm up" become their real workflow
# Your PR structure:
skills/
your-skill-name/
SKILL.md
examples/
README.mdThank You
Go build something useful.
Next event coming soon · Stay tuned in the NLM community