Workshop 2.0

VIBE CODING

From prompting to designing the flow.

When
April 2026
Where
NYU Tandon
Format
Hands-On
Your Guides

Who's running this.

Poe
Poe
Founder · Engineer
  • nyantrace.ai · MOT 2026, NYU
  • Ex-VP Engineering, 99.co
  • Ex-Eng Lead, Shield (anti-fraud ML)
Sally Yang
Sally Yang
AI Product Thinker
  • MOT 2026, NYU
  • Ex-CoS, TuSimple
  • Co-host · Moderator
The Shift

From writing prompts
to designing systems.

1.0
Individual Contributor
  • You type prompts.
  • AI writes scripts.
  • Each session starts fresh.
  • Memory lives in your head.
2.0
Systems Architect
  • You design environments.
  • Agents work autonomously.
  • Steps share state.
  • Memory lives in the system.
Agenda

What we'll cover

01
Workflow

What chains of steps look like.

02
Memory

How agents remember anything.

03
Permissions

What an agent is allowed to touch.

Then we build one in Gumloop, together. Drag-and-drop.
01 · Workflow

A chain where each step's output
becomes the next step's input.

Trigger
Step
Step
Output

Each box does one thing. The whole becomes the system.

02 · Memory
AI agents forget.
02 · Memory
Write it down.

If a step needs to know it tomorrow, store it today.

Files

Daily notes. State files. Logs.

Variables

Gumloop carries values between nodes.

Databases

Notion pages. Sheets. SQL.

03 · Permissions

Don't hand out
house keys.

🔑
Key
Full access

Anything you can do, the agent can do.

📬
Box
Mailbox only

Read what comes in. Nothing else.

03 · Permissions

Give the minimum it needs.

Gmail
ReadALLOW
SendDENY
Slack
PostALLOW
DeleteDENY
Notion
AppendALLOW
OverwriteDENY

Why? Safety + cost + predictability.

The Pattern

Every workflow is the
same three things.

01
Workflow
The chain of steps.
02
Memory
What's stored between steps.
03
Permissions
What each step can touch.
Now let's build one.
Anatomy

What makes a workflow good.

01
Small steps

One job per node. Easier to debug, easier to swap.

02
AI only where judgment is needed

Glue with code. Reason with AI. You don't need an agent for everything.

03
Memory between steps

Store state outside the prompt. Files, variables, databases.

04
A human checkpoint

At the moments that matter. Especially before something irreversible.

Real Example

My vault is a workflow.

Folders are the system.

Vault/
├─ CLAUDE.md
├─ Inbox.md
├─ Task Board.md
├─ Scratch Pad.md
├─ Daily Notes/
├─ Agent Daily Notes/
├─ Sessions/
└─ Research/
Workflow
Morning → Evening loop
Daily Notes + Agent Daily Notes form the chain.
Memory
CLAUDE.md + Daily Notes
Rules persist in CLAUDE.md. State persists in dated files.
Permissions
Whitelisted tools
Agents can Read / Write / Edit — nothing destructive.
What We're Building

A job application assistant.

Twelve nodes. Three of them think.

01
🔍
Search Google
Forward Deployment Engineer listings.
02
🌐
Scrape Greenhouse
Anthropic's job board.
03
🔗
Combine
Merge into one list.
04
📄
Read Resume
From Google Docs.
08
🎯
Filter ≥ 7
Drop weak matches.
AI
07
🧮
Extract Score
Parse score + reasoning.
AI
06
🤖
Score Jobs
Rate 1–10 vs resume.
05
📋
Duplicate × N
One copy per job, in parallel.
09
📋
Duplicate × N
One copy per qualifying job.
AI
10
✍️
Tailor Resume
Rewrite per role.
11
📝
Write Doc
New Google Doc per job.
12
📊
Log to Sheets
Tracker spreadsheet.
3 AI nodes — score, extract, tailor 9 code nodes — fetch, duplicate, filter, write Permissions: read web · write Docs & Sheets only
Workflow Builder

One prompt. Twelve nodes.

Describe the flow. Gumloop builds it.

PASTE INTO GUMLOOP Workflow Builder prompt
Build me an automated job hunting flow that does the following:

 Search for jobs using a Google web search targeting site:boards.greenhouse.io
  for a specific role (e.g. "Forward Deployed Engineer" or "Forward Deployment Engineer")

 Scrape a specific company's job board (e.g. Anthropic's Greenhouse page at
  boards.greenhouse.io/anthropic) to get their latest listings

 Combine both job sources into unified parallel lists of:
  job URLs, job descriptions, company names, and job titles

 Read my resume from a Google Doc link I provide

 Score each job using AI — compare my resume against every job description
  and generate a match score from 1–10 with a 1-2 sentence justification

 Filter jobs that score below a minimum threshold (e.g. 7 out of 10)

 Tailor my resume for each qualifying job using AI — rewrite it to mirror
  the job's language and highlight the most relevant experience, without inventing anything

 Save each tailored resume as a separate Google Doc

 Log all results into a Google Sheets job tracker with these columns:
  Title, Company, Job URL, Score, Justification, Resume Link
Live Build

Let's open Gumloop.

Watch for three things.

the chain
the memory
the permissions
Surfaces

Once you see the pattern, you can build it anywhere.

01
Gumloop
Visual nodes. Best for learning.
02
n8n / Zapier
Visual + code. Production glue.
03
Claude Code
Agents in your terminal. Tool whitelisting.
04
Obsidian + scripts
Files as memory. Cron-driven.
You are a vibe coder.
You're a supervisor
of intelligence.
QUESTIONS?
WhatsApp group QR
Stay in the loop
Join the chat.
chat vibecoding · WhatsApp
VIBE CODING 2.0 · BUILT FOR HUMANS
SLIDE 01/17