Live Workshop • April 16, 2026

From Prompt to Production
in One Session

We built an AI-powered LinkedIn Chrome extension — live, from a voice memo, using multi-agent orchestration.

Claude Code + OMC Qwen 3.5 (35B) Zero lines typed by hand

What We Built

A Chrome extension that crawls your LinkedIn feed using Qwen 3.5 via OpenRouter to classify connections and build actionable dossiers for specific networking goals.

💜

Nonprofit Support

Find donors and allies for a friend's nonprofit focused on AI for human good.

💰

Fundraising

Surface investors and angels for Multpod, an AI event management startup.

🤝

Design Partners

Event organizers, conference ops, and community leads for Multpod.

📍

Bay Area Networking

Location-aware lists for Palo Alto, East Bay, and SF meetups.

Tools We Used

♦️
Claude Code
Anthropic's AI coding CLI
⚙️
oh-my-claudecode
Multi-agent orchestration
Qwen 3.5 (35B)
Post classification via OpenRouter
🛠️
Chrome Extension
Manifest V3 + side panel

Five Phases, One Session

💬
Interview
/deep-interview
🗺
Plan
/ralplan
🛠
Architect
Review
Build
/autopilot
Test
Smoke test

How It Went Down

Setup

Install Claude Code + OMC

Installed Anthropic's CLI, then added oh-my-claudecode for multi-agent superpowers.

Deep Interview

Refine the raw idea

OMC's Socratic interview turned a rambling voice note into a tight spec with goals, constraints, and edge cases.

Ralplan

Plan → Architect → Review

Three agents reached consensus on the implementation plan before a single line of code was written.

Autopilot

Autonomous build

Designer, executor, and verifier agents built the full extension end-to-end.

Smoke Test

Load & verify in Chrome

Extension loaded, side panel opened, LinkedIn feed crawled, posts classified. It works.

Where It All Began

I want a Chrome extension that uses Qwen 3.5 on OpenRouter to scroll through my LinkedIn, classify posts, and build dossiers of people I should meet — for my startup fundraising, a friend's AI-for-good nonprofit, and finding design partners. Do it as a Chrome extension so I can watch. Don't get me banned.

How It Works

LinkedIn Feed
Content Script
Background Worker
Qwen 3.5
↓ classify & score ↓
IndexedDB
Side Panel
Lists & Messages

Why This Approach Wins

The Old Way

Prompt → code → debug → repeat. Hours of back-and-forth fixing what the AI missed.

The OMC Way

Interview → plan → architect → build → verify. Ambiguity dies before code starts. One shot, working extension.

Session Stats

12
Agents Spawned
154
LLM Calls
13.4M
Tokens
$34
Total Cost
0
Lines by Hand
163
Tool Calls
7
Agent Types
2h
Build Time

Agents Used

Planner ×2 Architect ×2 Critic ×2 QA Tester ×3 Explorer ×1 Code Reviewer ×1 Scientist ×1

97% of tokens were cache reads — Claude Code reuses context aggressively. Actual LLM generation: 111K tokens. Cost breakdown: cache reads $19, output $8, cache writes $6.

Remember This

1 Talk, don't type specs

The deep-interview turns your rambling into a real specification. Just answer honestly.

2 Plan before you build

10 minutes of /ralplan saves hours of debugging. Let agents argue about architecture first.

3 Specialists beat generalists

Designer, executor, reviewer, verifier — each does one thing well. That's the whole trick.

4 Open models ship

Qwen 3.5 at 35B handled classification perfectly. Right model, right cost, right results.

5 The future is agentic

Voice memo → working Chrome extension. In one live session. This is software now.

Get Started

# Install Claude Code
$ npm install -g @anthropic-ai/claude-code

# Install oh-my-claudecode
$ git clone https://github.com/yeachan-heo/oh-my-claudecode
$ cd oh-my-claudecode && ./install.sh

# Go
$ claude
> /deep-interview
> /ralplan
> /autopilot
View on GitHub →
🦞🦞🦞🦞🦞
You made it to the lobsters. You're one of us now.
🦞 🦞 🦞