Y CombinatorConductor CEO Charlie Holtz Walks Us Through His AI Coding Setup
At a glance
WHAT IT’S REALLY ABOUT
Conductor CEO demos agentic AI workflow, guardrails, and beliefs
- Holtz demonstrates a multi-workspace, keyboard-driven workflow where he continuously spawns AI tasks, reviews diffs, leaves PR-style comments, and merges from inside Conductor.
- He argues that voice input and a visual GUI are increasingly important because humans manage parallel agent work better spatially than through a terminal alone.
- Conductor is built with strong process constraints—workspaces map to worktrees and must produce PRs—so AI changes are reviewable and the codebase stays controllable.
- He emphasizes quality boundaries like “slop-free zones,” plus the principle “don’t let the AI be your architect,” keeping core APIs/contracts and UI decisions human-led.
- Holtz compares models tactically (Codex as the tool-calling workhorse vs Claude/Opus as the creative partner) and shares realities of token spend, fast-mode usage, and evolving workflows away from IDE/GitHub.
IDEAS WORTH REMEMBERING
5 ideasTreat AI agents like parallel teams you supervise, not a single assistant.
Holtz runs many workspaces at once, bouncing between them while agents execute; the human role becomes prioritizing, reviewing, and steering rather than typing code.
Workflow constraints are a feature when AI writes most code.
Conductor forces changes to happen in isolated worktrees that become PRs and require merging, creating natural checkpoints and preventing silent, unreviewed edits.
Keep “slop-free zones” to stop quality from compounding downward.
He maintains parts of the code/docs that are known-human-written (or strictly human-reviewed) so models don’t train on and amplify earlier low-quality AI output.
Don’t delegate architecture and product craft to the model.
Holtz says humans must own core abstractions (e.g., the workspace concept), API/contracts, and UI interaction decisions; AI can iterate inside safer, non-core areas.
Use customization to encode engineering culture, not endless knobs.
Their Claude.md/skills files include explicit guidance like “we’re a startup, not enterprise,” shaping outputs toward team norms without relying on heavy analytics or A/B tests.
WORDS WORTH SAVING
5 quotesWe are all trying to talk to our computers more.
— Charlie Holtz
Most of the time if I want small edits, I'll, like, highlight and then, uh, tell the AI, um, about my comments, or I'll just, like, speak into my computer and say, "That button looks a little too wide. Like, can you, uh, can you make it smaller?"
— Charlie Holtz
But the ideal is, like, you should feel like the CEO of a little company, and you can see all your agents working for you, and they'll bring you up, like, digestible reports, and then you can point them in the right direction if they, like, need some correction or just merge it in if it looks good.
— Charlie Holtz
I think something that's really important to us is having, like, clear boundaries between, uh, well, we call them slop-free zones, um, and having, like, parts of the code base or, like, parts of the documentation that we, like, know is written by a human.
— Charlie Holtz
Code is almost like, uh, sawdust now in that, like, it used to be that code was the thing you were building. It was, like, the structure. You were putting time into, like, in, in, into, like, crafting the code, and now you're putting time into describing what you want and how you want it to be built, and the code is almost just, like, sawdust that comes out of that process.
— Charlie Holtz
High quality AI-generated summary created from speaker-labeled transcript.
