Vibe Code Camp Distilled

Nat Eliason - Building with Claude Bot and AI-First Engineering Management

Nat Eliason - Building with Claude Bot and AI-First Engineering Management

Key Insights

Summary

Nat Eliason is an author, entrepreneur, and self-described “inveterate vibe coder” who has reinvented himself multiple times - from running Growth Machine (a content marketing agency) to writing about crypto, to now building AI-first products. In this presentation, he demonstrates his complete workflow for building Tegan, an AI-first content marketing platform that recreates what his agency did but with autonomous agents instead of human contractors.

The demo showcases a fundamentally new approach to software development where the builder becomes more of a director than a hands-on engineer. Nat uses Conductor (a GUI wrapper for Claude Code) to manage multiple parallel development streams via git worktrees, employs custom GitHub workflows that automatically review, fix, and iterate on PRs, and runs Claude Bot on a dedicated Mac Mini as an always-on engineering manager. Claude Bot monitors Sentry errors in Slack, reviews user conversations nightly, identifies bugs and UX issues, and orchestrates Codex sub-agents to implement fixes - all autonomously.

What makes this particularly compelling is Nat’s honesty about the limitations and his systematic approach to solving them. Rather than claiming AI can do everything, he’s built a sophisticated orchestration layer where Opus 4.5 acts as an engineering manager reviewing Codex’s implementation work, where automated workflows narrow scope rather than expand it, and where he maintains creative control while delegating the implementation entirely. This isn’t theoretical - he’s shipping a production SaaS product this way.

Main Topics

From Vibe Coding to Production-Ready Development

Nat traces his journey from early 2024 when he started using Cursor with Claude Sonnet 3.5 (which “wasn’t very good”) to the present where he feels confident building production applications. The turning point came in late 2024 with the launch of Claude 3.7 and 4.0 when agentic modes became available.

“I launched a course called build your own apps uh teaching other people how to vibe code um unfortunately missed the term vibe coding by like a month” (00:02:35-00:02:42)

The critical shift happened with Opus 4.5 and Codex 5.2:

“really in the the last two months three months since opus 4.5 and codex 5.2 launched um there there’s been a we’ve like we’ve crossed that threshold where it really does feel like you can just go all the way with this stuff now” (00:03:14-00:03:20)

Nat emphasizes that even senior engineers are now publicly validating this approach - DHH (creator of Ruby on Rails) saying “this is really good now” carries more weight than hobby engineers making similar claims.

Tools mentioned: Cursor, Claude Sonnet 3.5, Claude Opus 4.5, Codex 5.2, Replit

The Mac Mini as 24/7 Development Infrastructure

One of the most striking revelations is Nat’s decision to buy a dedicated Mac Mini specifically to run Claude Bot continuously:

“i literally went out on on monday and bought a separate mac mini to set up in my office so i can have claude bot running 24 7. uh and i’m literally like sending it requests in bed at 10 p.m like hey claude i’m going to sleep but like why don’t you start on this and just get it as far as you possibly can” (00:04:20-00:04:36)

He frames this as equivalent to hiring an engineer:

“it’s like oh this 500 device with a 200 a month subscription is literally like hiring another engineer and and maybe you know and in some ways better in some ways worse” (00:05:59-00:06:11)

Nat notes he now has two $200/month Claude subscriptions because the productivity gain justifies it. The unintended consequence: Mac Minis becoming essential AI development infrastructure.

Conductor: The GUI Layer for Claude Code

Nat demonstrates Conductor Build, which he describes as “basically a wrapper around claude code and codex.” Key advantages over terminal-based Claude Code:

“this feels like or in the stone age this is like two months ago like why would you want right like i have to i have to type like open a pr right or i have to like use the i have to like do the slash commands for skills it’s just like i don’t know i mean it’s fine it works but like conductor is so nice” (00:11:10-00:11:29)

The terminal vs GUI debate: Nat argues forcefully that GUIs are superior for this workflow, acknowledging the only reason to use terminal is “to like look cool in a coffee shop.” The key insight is managing parallel workstreams - Conductor automatically creates git worktrees so multiple development threads don’t conflict until PR time.

Model orchestration within Conductor: Nat uses Codex 5.2 for initial implementation, then opens a new tab with Opus 4.5 in the same git worktree to review Codex’s work or write up PRs. This multi-model approach is “just like a pain in the ass to do in the terminal.”

Autonomous PR Review and Implementation Workflows

Nat built custom GitHub workflows that extend Claude’s built-in PR review capabilities to create a semi-autonomous loop:

  1. When a PR is opened, Claude reviews it automatically
  2. Claude reviews its own review and categorizes feedback:
  3. Issues to address in this PR
  4. Issues to create as separate follow-up issues
  5. Claude creates those follow-up GitHub issues
  6. Claude implements the “must fix now” items
  7. When the PR is closed, Claude evaluates the follow-up issues
  8. For issues deemed important, Claude implements them and opens new PRs

“literally all i’m doing now is opening prs going and making sure the like stuff ran properly and then after 10-15 minutes all the follow-up work has been done and i can just merge it and move on to the next thing” (00:15:51-00:16:03)

The key insight here is scope narrowing rather than scope expansion:

“it’s scope narrowing it’s saying like just fix the things that have to be fixed and then create these follow-up issues and then it can look at those follow-up issues and it can say like we don’t actually need to do this this is scope creep we’re not going to address that” (00:16:20-00:16:33)

Nat also has workflows that automatically fix lint and test failures.

Claude Bot: Always-On Engineering Manager

Claude Bot is the centerpiece of Nat’s workflow - an always-running instance of Claude Code that can orchestrate other services and respond to triggers from multiple sources.

What is Claude Bot:

“think about claude bot as like an always running claude code instance on your computer that can run any other service on your computer that you could run through the command line so you’re kind of like giving it full computer control” (00:20:52-00:21:06)

Capabilities demonstrated: - Email management (finding, drafting, canceling subscriptions) - Browser automation via Chrome - Telegram integration for conversational tasking - Slack integration for monitoring and responding - GitHub management and code implementation - Custom API integrations

Primary interface - Telegram:

“i primarily talk to it through telegram… i asked it like hey we’re going to do a live demo of how we work together what should we show off to get people excited and it came up with some ideas” (00:17:20-00:17:39)

Nat can send voice notes while walking to a coffee shop, and Claude Bot starts working on requests immediately.

Opus as Orchestrator, Codex as Implementation

A critical architectural decision is using Opus 4.5 to orchestrate Codex 5.2 instances rather than doing all the work with one model:

“i’ve designed it so that it has access to uh spin off codex sub-agents so when it needs to do a little bit more of a coding lift it actually creates sub-agents using codex instead of just clawed normal sub-agents” (00:19:09-00:19:21)

Two reasons for this approach:

  1. Quality: “codex does a little bit better with coding than opus 4.5 it like goes a little bit slower digs a little deeper and builds out slightly better solutions but opus is a much better like general intelligence model so i trust opus to be my engineering manager” (00:19:27-00:19:43)

  2. Rate limits: “i’m not using up all my claude credits on like everything so claude is like the orchestrator and then it’s using codex credits for a lot of the programming” (00:19:48-00:19:54)

The workflow: Codex sub-agents implement and fix issues, then report back to the Opus instance which does a final review before merging PRs.

Automated User Feedback Analysis

One of the most sophisticated automations Nat built: Claude Bot analyzes user conversations nightly to identify issues and improvements.

The workflow: 1. Every night at 2 AM, Claude Bot pulls all user conversation logs from Tegan’s servers 2. It reads through conversations looking for: - User frustrations - Weird behavior - Things that failed 3. It writes a report on everything that happened and ways to improve 4. Bugs get fixed automatically 5. Improvement suggestions get sent to Nat in a Google Doc (because he prefers reading there) 6. Nat sends back a voice note approving improvements 7. Claude Bot creates a plan 8. Once agreed, it kicks off Codex agents to implement

“that’s like a whole new level from me having to watch every single conversation and me having to figure out you know what should i tell claude code to do it’s like that’s kind of abstracted now because opus is a better engineering manager than i am i trust its judgment on a lot of these things” (00:25:56-00:26:12)

Sentry Integration for Automatic Bug Fixing

Nat has Claude Bot integrated with Sentry (error monitoring) via Slack:

“i also have claude bot looped into slack where anytime there’s a bug on tegan century sends that bug to this channel in slack and it tags claude bot and then you can see claude bot receives the error it you know looks into it and then it automatically spins up a codex sub-agent to code up a fix and then open a pr” (00:22:51-00:23:06)

When bugs are straightforward, Claude Bot fixes them autonomously. When they’re complex, it asks clarifying questions in Slack before implementing.

Browser Automation for Testing

Nat built a full testing suite where Claude Bot: 1. Opens the staging deployment in a browser 2. Creates a test account 3. Goes through the onboarding flow 4. Looks for issues and weird behavior 5. Writes up a report 6. Automatically implements fixes for any bugs found

He runs this overnight to avoid using daytime credits.

The Death of Typing, CLIs, and Manual Review

Throughout the demo, Nat emphasizes that multiple traditional development practices are now obsolete:

Typing:

“obviously i’m using monologue for my voice to text because it’s a wonderful tool and uh typing is also not something we’re bringing into 2026 nobody should be typing anymore you should just be using voice of text because it’s way better typing is over typing’s done” (00:18:10-00:18:21)

CLIs vs GUIs:

“the cli is the stone age from two months ago and guis are back uh the rest of the world will probably figure that out in a month or two” (00:13:31-00:13:43)

Manual code review: With autonomous PR review, implementation, and testing, the developer’s role shifts from reviewer to director.

Tegan: The AI-First Content Marketing Platform

The product Nat is building (Tegan, named after Chrissy Teigen based on an AI-generated friendly marketer face) recreates his Growth Machine agency as an AI-first solution.

Design approach:

“i i so there’s a a claude skill called front end design it’s one of the most popular claude skills i highly recommend installing it if you haven’t and then i picked a primary color which is the like fallow green… i literally just said like hey claude use your front end ui skill and create a landing page and use this green as the primary color and go and it came up with this” (00:08:37-00:09:07)

The app is in soft launch - Nat created a coupon code “every” for a free month, acknowledging “it’s not completely ready there’s still a lot of stuff that i’m fixing and like you’re going to hit bugs.”

URL: hiretegan.com

Reinvention and Commitment

When asked about his pattern of reinventing himself (marketing agency → crypto author → vibe coder), Nat gives an unexpectedly honest answer:

“i wouldn’t actually like recommend it as a career path because like like yes it it has worked out well for me but most of the time if you’re this like non-committal to stuff it like doesn’t go well for you and i’m actually trying to move away from that you know i like i i do switch too often i do jump to new things too often like i don’t stick with stuff for it to really hit the the compounding gains that it could have” (00:28:04-00:28:29)

With Tegan, he’s putting “everything else on hold for the most part and i’m just all in on mastering this new way of building and applying it to building something where i have a very strong competitive knowledge advantage.”

Actionable Details

Tools & Products

Conductor Workflow

  1. Open Conductor Build
  2. Select repository
  3. Click to create new workspace (auto-creates git worktree)
  4. Give Codex 5.2 initial implementation prompt
  5. Open new tab with Opus 4.5 in same worktree for review
  6. One-click open PR
  7. One-click spin up dev server to preview changes

GitHub Workflow Setup

Nat extended Claude’s built-in GitHub workflows to add: - Self-review of initial review - Automatic categorization of issues (fix now vs. create issue) - Follow-up issue creation - Automatic implementation of “must fix” items - Issue evaluation on PR close - Recursive PR opening for important follow-ups - Automatic lint/test failure fixes

To set up the basic version: Use Claude Code’s /github install command, then ask Claude to extend it.

Claude Bot Setup for Multi-Model Orchestration

  1. Install Claude Bot on a dedicated machine (Nat uses Mac Mini)
  2. Configure access to multiple AI models (Codex, etc.)
  3. Set up Telegram integration for conversational interface
  4. Connect to Slack for monitoring channels
  5. Configure custom skills:
  6. Sentry error monitoring
  7. User conversation analysis
  8. Browser automation testing
  9. Email integration
  10. Create scheduled tasks (e.g., 2 AM nightly user conversation review)

User Feedback Loop

  1. Collect user conversation logs from application servers
  2. Create custom API for Claude Bot to access logs
  3. Schedule nightly pull (2 AM)
  4. Claude Bot analyzes for:
  5. User frustrations
  6. Weird behavior
  7. Failures
  8. Generates report in Google Doc
  9. Sends to developer for approval via voice note
  10. Creates implementation plan
  11. Spawns Codex agents to implement approved changes

Voice-Based Development Workflow

  1. Use Monologue or similar voice-to-text app
  2. Send voice notes to Claude Bot via Telegram
  3. Claude Bot transcribes and processes requests
  4. Responds with status updates
  5. Implements while you do other things
  6. Review results when convenient

Resources Mentioned

Quotes Worth Saving

“i literally went out on on monday and bought a separate mac mini to set up in my office so i can have claude bot running 24 7. uh and i’m literally like sending it requests in bed at 10 p.m like hey claude i’m going to sleep but like why don’t you start on this and just get it as far as you possibly can uh and then like you know and i’ll show you guys how i’m doing this and then i just like wake up and i’ve got like a whole new feature built and it’s been tested and it’s i mean it’s magical” (00:04:20-00:04:47)

“we’re almost like speed running the dos era… 15 years to get to guis like it’s just we we had dos for two months and now we’re into guis you know” (00:11:49-00:12:01)

“opus is a much better like general intelligence model so i trust opus to be my engineering manager and to control all of my codex instances to like write the code” (00:19:34-00:19:43)

“that’s like a whole new level from me having to watch every single conversation and me having to figure out you know what should i tell claude code to do it’s like that’s kind of abstracted now because opus is a better engineering manager than i am i trust its judgment on a lot of these things especially as we talk about it” (00:25:56-00:26:12)

“typing is also not something we’re bringing into 2026 nobody should be typing anymore you should just be using voice of text because it’s way better typing is over typing’s done” (00:18:10-00:18:21)

“the like what’s next is moving from uh having a really competent engineer at your disposal to having a really competent engineering manager at your disposal” (00:24:29-00:24:40)

← Back to Index ← Katie Parrott Tina He →