Vibe Code Camp Distilled

Kevin Rose and Kieran Klaassen on Compound Engineering

Kevin Rose and Kieran Klaassen on Compound Engineering

Key Insights

Summary

Kevin Rose (founder of Digg, serial entrepreneur, and investor) and Kieran Klaassen (creator of the Compound Engineering plugin and GM of Quora at Every) discuss the evolution of AI-assisted software development and the philosophy behind Compound Engineering. Kevin shares how he went from struggling with context window management and manually orchestrating multi-agent workflows to achieving 97% first-compile code quality using Kieran’s plugin.

Kieran explains how Compound Engineering emerged from building Quora (an AI email agent) for 18 months, distilling traditional engineering practices—planning, working, reviewing, and compounding—into a structured workflow for AI coding. The conversation covers practical details like using DeepWiki to understand codebases, converting them to NotebookLM podcasts for faster learning, and running Mac Minis 24/7 to continuously ship features. They also discuss the philosophical shift happening in engineering: ICs becoming tech leads, solo developers managing manufacturing-style processes, and teams of one becoming viable.

Main Topics

The Evolution from Vibe Coding to Production Engineering

Kevin describes his journey over the past eight months, starting with disappointing results and gradually developing his own multi-agent workflow. He would: - Use deep thinking modes to rant for 10-15 minutes about what he wanted to build - Front-load models with massive context about tech stack (Next.js, Vercel, Supabase) - Run two side-by-side panes in Cursor: Opus 4.5 for heavy lifting, GPT-5 as a “friend” reviewing the code - Manually pass feedback back and forth, forcing models to justify design decisions

“I was manually doing these multi-step, multi-agent process back and forth and back and forth. And I was taking my code quality from about, you know, call it like first compile, first build, working to, you know, from call it 80% to like maybe close to 90% or something like that. But it was very laborious.” [00:09:21]

This manual process worked but wasn’t scalable. When friends heard the workflow, they said it was too much work.

Why Planning Matters (And Why We Forgot)

Kieran emphasizes that planning is fundamental to all engineering:

“If you think about normal engineering work, like the first thing you do is you go plan. And for some reason, when we started AI coding, we just kind of forgot to plan. It was like one shot everything.” [00:13:25]

The plugin’s planning workflow mirrors how capable teams actually work. Kevin notes:

“I feel that as I’ve coded with these models over time, like they’re being trained on high quality code and they’re being trained on shitty code. And so in some sense, you’re kind of getting some type of average of it all… if you plan like a senior engineer, in some sense, it almost has to like call those engineers to the table to go and actually execute against that work.” [00:15:43]

Kieran agrees this aligns with recent Anthropic research showing models respond differently based on the user’s knowledge level.

The Compounding Secret Sauce

The compounding step is what differentiates this workflow from simple plan-execute loops. Kevin asks for technical details:

Kevin: “The compounding part is like, you kind of don’t know what’s going on. You see it working, but technically what is happening there?” [00:20:12]

Kieran explains: - Run /workflows compound when something goes wrong or you want to capture a lesson - It extracts what needs to be learned and creates documentation in a docs folder - Files are markdown with YAML frontmatter containing keywords for search - Instead of bloating context windows with everything, it uses genetic search to find relevant docs - A best practices agent (using Haiku for speed/cost) searches and injects only applicable lessons into planning and review - Kevin asks: “How do you know when to load them in?” - Kieran: “I just have them baked into the planning and the review… that agent will start to do, like a genetic search, finding all the relevant files” [00:21:42]

This creates a custom knowledge base—like Context7 but for your own code patterns.

Kevin’s Research Workflow

Kevin shares a fascinating meta-learning workflow:

  1. Paste the Compound Engineering GitHub repo into DeepWiki (from Devin)
  2. Get a full technical brief of how it works
  3. Dump that into NotebookLM to generate a 15-minute podcast
  4. Listen at 2x speed (7 minutes total)
  5. Decide if it’s worth deeper investigation

“There’s so much coming at us right now, like so many new products and features. How can we then take what multiple people are saying we should go spend hours on and compress it down into something that we can quickly grok and understand.” [00:11:18]

The DHH Reviewer

The plugin includes reviewer personas, including one styled after DHH (David Heinemeier Hansson). Kevin sent DHH his own AI-generated code review, and DHH confirmed “that’s actually something I would have said.” [00:29:17]

Some reviews reportedly say things like “Just delete the whole plan. This is crap.”

“There’s something about like, I know, like, I love how opinionated he is. Like that’s the whole thing about, you know, 37 signals and Jason Fried and DHH is like, you got to respect, like they do it their way… no, no, no, this is the way we do it.” [00:28:47]

Kieran: “That’s also kind of the spirit of the compound flow is like just having strong opinions about how to do software because otherwise it’s just getting mediocre” [00:29:04]

End-to-End Autonomous Coding with /lfg

Kieran reveals the newest command: /lfg (Let’s F’ing Go)

“I introduced a new command called slash LFG. And basically that’s just capturing all of those things in a relv loop and what it does, it does everything. So it’s just like, you have an idea until PR is up with screenshots and videos and like tested in a browser.” [00:24:54]

This represents a shift from manual orchestration to fully autonomous workflows. With Opus 4.5, you can go from idea to PR without human intervention.

Kevin’s response: “I literally just bought a Mac mini to run 24/7” for continuous feature development.

Kieran notes he uses Mac minis for automation because they integrate with iMessage via AppleScript, making notifications easy without complex infrastructure.

The Vibe Coding to Production Pipeline

Both Kevin and Kieran use V0 (Vercel’s AI design tool) for rapid prototyping before production work:

Kevin: “For me, a product has to feel right before I want to build it. So even though I kind of have to like, I’m a very visual person… I just crank on V zero, like, you know, I just let that go in multiple windows with multiple projects. And then when I get that gut feeling like this needs to exist, then I jump over and I’ll just download that, throw it into its own folder and then let compound engineering go to town on it” [00:26:32]

Kieran: “I’m doing exactly the same, which is cool. I call mine baby Cora, which is just my mockup version… I just vibe code whatever I want.” [00:27:01]

This workflow: Vibe code in V0 → Get design right → Export → Run through Compound Engineering for production implementation.

What’s Next for Compound Engineering

Kieran outlines upcoming directions:

  1. Discovery phase before planning: More brainstorming tools (new /brainstorm command just launched)
  2. Multi-platform support: Now available on OpenCode and Codex in addition to Claude Code
  3. Design iteration: Pushing more toward the design and iterative sides
  4. Model experimentation: Testing different models beyond Claude

Kevin shipped tens of thousands of installs in just a few weeks, with contributors now submitting PRs. One user ran compound as an “initializer” to extract everything from a repo and seed the documentation automatically.

Aphantasia and Coding

Kevin shares a personal revelation:

“I didn’t realize actually until a few weeks ago that I have this thing called aphantasia, which is a condition where people cannot visualize something in their mind’s eye… I was always wondering, why can I just not remember syntax and certain things the way that my friends can? They were just so much faster than I was.” [00:06:16]

This pushed him toward product design rather than pure development. Now, AI tools have enabled him to “kind of do it all” without needing to memorize syntax.

Actionable Details

Tools & Products Mentioned

Kevin’s Pre-Compound Workflow

  1. Hold function key, use Whisper flow to rant for 10-15 minutes
  2. Specify tech stack: Next.js, Vercel, Supabase
  3. Get massive planning document back
  4. Two Cursor panes: Opus 4.5 (builder) + GPT-5 (reviewer)
  5. Ask GPT-5: “How’d I do? Check it out.”
  6. Pass feedback to Opus: “My other friend looked at the code and said they would have changed these things. Don’t agree with me. Don’t just wholesale go in and make changes. Tell me why you’re about to make those changes.”
  7. Iterate manually between models

Result: 80% → 90% code quality, but too laborious

Compound Engineering Workflow

Basic usage: - /workflows plan - Create structured plan - /workflows work - Execute the plan - /workflows review - Multi-agent security, performance, best practices review - /workflows compound - Extract lessons, create documentation

Advanced: - /lfg - Autonomous end-to-end from idea to PR - /brainstorm - Pre-planning ideation - create a plan - Natural language invocation (plugin stays loaded in context)

No need to re-invoke: Once the plugin is loaded, Claude understands the workflow. You can just say “create a plan” or continue conversations without the slash command.

The Compounding Process

  1. Something goes wrong or you want to capture a pattern
  2. Run /workflows compound
  3. Plugin extracts lessons from current context
  4. Creates markdown file in docs/ folder with:
  5. YAML frontmatter with keywords
  6. Specific patterns to follow/avoid
  7. Context-specific guidance
  8. During future planning/review, best practices agent:
  9. Does genetic search on docs folder
  10. Uses Haiku (fast, cheap) to find relevant files
  11. Injects only applicable lessons into context

Benefit: Thousands of documented patterns without bloating context windows

Learning New Tools Fast (Kevin’s Method)

  1. Find GitHub repo
  2. Paste into DeepWiki → get technical brief
  3. Put brief into NotebookLM → generate podcast
  4. Listen at 2x (7 minutes for 15-minute podcast)
  5. Decide if worth deeper dive

Quotes Worth Saving

“I will say that, yeah, I’m in agreeance with you in that this world of kind of vibe coding is transitioning from vibe to actually just coding. Right? And, you know, when you see Linus Torvalds commit something via Claude or whatever he was using, I think it’s safe to say this is kind of coding now.” [00:05:04] — Kevin Rose

“For some reason, when we started AI coding, we just kind of forgot to plan. It was like one to one shot everything. We’re like, Oh, I want this feature. Do it. And we complain and I’m like, yeah, but if you just think about it, like how would you do this if you had a team of capable people, it’s just, you go plan.” [00:13:25] — Kieran Klaassen

“If you are very knowledgeable, like it will respond differently to you than if you are not. So yes. And then on top of that, like the grounding you do in these flows is like, you’re pulling the things in that you want… You’re pulling the research in like your style and you just surround it with all the stuff you wanted to do good with.” [00:16:51] — Kieran Klaassen on how planning quality affects model output

“I introduced a new command called slash LFG. And basically that’s just capturing all of those things in a relv loop and what it does, it does everything. So it’s just like, you have an idea until PR is up with screenshots and videos and like tested in a browser.” [00:24:54] — Kieran Klaassen on end-to-end autonomous coding

“That’s also kind of the spirit of the compound flow is like just having strong opinions about how to do software because otherwise it’s just getting mediocre and it doesn’t mean you have to agree with DHHs all the time, but like sometimes the DHH reviewer has a point and it makes you think like, Hey, maybe this is the complete wrong thing I’m building.” [00:29:04] — Kieran Klaassen on opinionated code review

← Back to Index ← Geoffrey Litt Thariq Shihipar →