Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup
In a Nutshell
Charlie Holtz built Conductor, an Electron app that orchestrates multiple Claude/Codex agents through workspaces, PR reviews, and a three-panel GUI. He enforces strict human-led architecture, "slot-free zones," and skills files while using voice commands, fast mode, and Telegram for remote control. The workflow treats code as disposable "sawdust," prioritizes minimal output, and keeps humans conducting rather than letting AI architect.
These notes were generated by AI and may contain inaccuracies.
Charlie is the co-founder of Conductor, an app that lets users orchestrate multiple coding agents on Mac. Conductor was part of YC Summer 24. A recent essential tool is a gooseneck microphone purchased for $20 on Amazon, used to facilitate speaking to computers more frequently.
In open floor plan offices, the microphone allows leaning over to whisper instructions to Claude, such as "Please merge PR 3475," reducing disruption compared to speaking aloud. Charlie spends most of his day in Conductor, using it to build Conductor itself. He frequently initiates new tasks using Command N, which provides a sneak peek of upcoming cloud workspaces functionality.
After pressing Command N, Charlie speaks commands such as "Can you take a look at the latest Linear issue and give me a rough pass at how you'd solve it?" and presses enter. The task appears running in the sidebar. While one agent works, he switches to another chat. Keyboard shortcuts are heavily utilized, including Command Shift Y to access a workspace ready to merge. He reviews the PR, and if it looks good, proceeds; if not, he adds GitHub-style comments for clarification.
A core aspect of Conductor usage is experimentation. Multiple workspaces are launched to test different ideas, with most not making it into production. Four PRs may be in review while numerous experimental ideas remain in progress. Successful experiments can be promoted to internal settings and then experimental settings.
Charlie demonstrates speaking into his phone to request a new feature for changing the theme to "hacker mode," then clicks "conduct" to initiate work on his computer remotely. When asked if he still writes code, he responds that he does so very occasionally, such as editing Tailwind classes or ENV files. A mode called caveman mode was added, allowing direct keyboard typing and file changes when needed. For small edits, he highlights text and provides verbal instructions to the AI.
Workspaces progress through states: "in progress" when kicked off, "in review" once a PR is created, and "done" once merged. A checks tab was recently added to monitor running checks. A dashboard page allows viewing all agents' work from one location and taking next actions. The vision is feeling like the CEO of a small company where agents deliver digestible reports for review or merging.
Telegram is used to communicate with agents. Spokenly handles text-to-speech via Control Space, running a local Parakeet model. The current machine has 128GB RAM to support local models. A MacBook Neo with the lowest RAM configuration was ordered to force usage of minimal specifications.
Significant time has been invested in skills files and Claude.md, which contains several hundred lines including statements on engineering practices: "We're a startup. You're probably used to writing enterprise code but that's not how we do things around here." Fast mode is always used for token efficiency. The context 7 MCP is used for documentation access. Claude runs with "dangerously accept all permissions," which is not the default.
Slot-free zones maintain clear boundaries where certain parts of the codebase and documentation are designated for human authorship. AI contributions to these zones require every line to be reviewed by a human. This prevents vicious cycles where AI sees bad code and generates more bad code. Specific lines in the codebase are marked "do not touch if you are an AI - this is for human eyes only."
Conductor is an Electron app using the native Safari web renderer with a Rust backend, though 90-95% of the desktop app is written in TypeScript. The web app is an Elixir Phoenix application, currently minimal since it only handles login. Charlie is a strong Elixir advocate and pushes for more Elixir usage when possible.
The principle "don't let the AI be your architect" is followed. The workspace concept in the sidebar was human-designed and will evolve. UI decisions, including the three-panel layout with chats on the left, chat interface in the middle, and review panel on the right, were carefully considered by humans. The "open in" button's icon placement in the top bar was debated extensively before being implemented.
Building core infrastructure around human-written APIs and contracts that AI contributes to less is recommended. Large portions of the codebase should allow AI free rein for experimentation without affecting core infrastructure. Current boundaries are described as murky and under improvement. Staying ahead of the frontier and pushing user comfort zones is a priority.
Initial feedback upon Conductor's launch was that managing three to five agents seemed overwhelming. The app deliberately prevents direct file editing, requiring all changes to go through work trees, PR creation, and merging. This enforces the intended workflow. The team constantly adapts to advancing model capabilities, moving toward agents that can run longer and smarter in unconstrained environments.
Conviction in product decisions comes from daily personal usage rather than analytics or A/B testing. If something doesn't feel right during use, changes are made quickly. The interface aims for a unified feel where messages can be typed without needing a separate composer.
Claude is used for back-and-forth creative work and new feature development due to its more collaborative, creative nature. Codex serves as the workhorse for specific problems requiring extensive tool calls and debugging. The choice depends on whether the task requires creative partnership or sustained execution.
Terminal interfaces are considered too restrictive for humans despite potentially suiting AI better. Spatial and visual cognition makes GUIs preferable. Conductor provides visual organization with chats on the left, review panels on the right, and chat interface in the middle. Many UI capabilities exceed terminal limitations.
Peak token spending reached $22,000 in July 2025 during early Conductor development, with tens of thousands of lines of code generated that month. Fast mode and high-effort settings are consistently used. However, minimal lines of code are prioritized to prevent codebase spiraling. The approach differs between starting new apps versus working in established codebases like Conductor.
Six months prior, hard PRs required opening an IDE for manual changes, and GitHub web app usage was frequent. Current workflow allows reviewing and commenting on code changes directly in Conductor. A checks tab integrates GitHub PR checks into the Conductor interface.
Notable user innovations include someone building a mobile version by spoofing IPC calls to the desktop app. Gary has demonstrated extensive Conductor usage, particularly with skills. A specific Gary mode was added that keeps all tool calls expanded by default and displays Gary's face.
Areas the team sees as underexplored include human-AI collaboration patterns such as communication with sub-agents and multiplayer chats where multiple humans work simultaneously with AIs. The conductor-orchestra metaphor describes the ideal workflow: conducting at the orchestra level while occasionally correcting individual instruments.
Code is increasingly viewed as "sawdust" - a byproduct rather than the primary focus. Time is spent describing desired outcomes rather than crafting code directly. Prompts become the critical asset that can be rerun with new models to generate fresh code. The submit-a-prompt feature represents an early experiment in malleable software.
Malleable software is compared to video games like Call of Duty, where the core structure remains consistent but users can customize elements like skins or reload speeds. The goal is enabling users to mod Conductor and build personal workflows while maintaining a crafted, consistent structure that feels thoughtfully designed.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
