<- all tokdocs

Headcount Organizes Claude Code Agents Like a Company, With Security and Legal as Blocking Reviewers

Watch on TikTok

View on TikTok ->

Headcount restructures Claude Code from a single prompt into a company org chart, where each department installs as its own plugin and agents divide work by what files they own rather than what topics they cover. The 23-second video walks through the open source repo cbrock84/headcount on GitHub, showing its README, file tree, and interactive org chart site. The narration hits three design decisions: per-department plugins so a project loads only what it needs, agent boundaries drawn around exclusive write surfaces, and a reviewer class where Security and Legal & Risk can block changes from every other department.

The Core Idea: Add a Department, Not a Prompt

The repo's tagline, shown in the opening frames, is "Add a department, not a prompt." Headcount models an agent organization for Claude Code as a company with a chief executive over 16 departments. The narration cites 143 skills, and the README shown on screen counts 172, so the catalog has grown as the repo evolved. Its current description reads "15+ departments, 125+ skills, each independently installable." Departments visible in the org chart include Office of the CEO (7 skills), Technology (19), IT Operations (12), Product (11), Marketing (19), and Demand Generation.

Each Department Is Its Own Plugin

Every department installs independently through the Claude Code plugin system. The on-screen setup takes two commands: /plugin marketplace add cbrock84/headcount to add the marketplace, then /plugin install security@headcount to install a specific department. The README argues that a smaller installed set triggers more sharply than loading everything at once, which is the "project loads only what it needs" point from the narration.

Skills Load Themselves by Request Matching

The Use section shown mid-video maps plain questions to the skill that loads. "Why isn't this landing page converting?" loads demand-generation:landing-page-cro-expert. "Review this design before we build it" loads security:threat-modeling. "Can we afford this hire?" loads finance:unit-economics. "Our growth has stalled" loads executive:business-growth-consultant. You can also invoke a skill directly by name, such as /finance:financial-modeling, when you want a specific lens. The department:skill naming convention prevents collisions across the catalog.

Agents Split by Exclusive Write Surfaces

The narration's sharpest point is how agent boundaries are drawn. Agents divide by exclusive write surfaces, meaning the files and outputs each one owns, instead of by topic. Because ownership is exclusive, two agents cannot quietly step on the same file. This addresses a real failure mode in multi-agent setups, where topically scoped agents overlap on shared code and silently overwrite each other's changes.

Security and Legal Sit Above as Reviewers

Two departments form a reviewer class that reports to the chief executive rather than into the functions they oversee. Security (CISO) and Legal & Risk review what the other departments commit to, and their blocking findings are not overrulable by the department under review. The README text shown on screen explains that this reporting structure exists precisely so a reviewer can stop work rather than add an opinion.

Traction Shown on Screen

The GitHub page captured in the video shows 959 stars, 148 forks, and a commit from one hour before recording, with MIT license and PRs welcome. The repo links to an interactive org chart at cbrock84.github.io/headcount where you can search all skills, filter to reviewer-class departments, and jump to source.

Key Takeaways

  • Headcount structures Claude Code as a company: a chief executive over 16 departments, each an independently installable plugin.
  • Installing only the departments you need keeps skill triggering sharp, since a smaller set matches requests more precisely.
  • Skills auto-load when a request matches their territory, or you can invoke one by name with the department:skill convention.
  • Agent boundaries follow exclusive write surfaces, so no two agents own the same files and silent overwrites are designed out.
  • Security and Legal & Risk are reviewer-class departments whose blocking findings cannot be overruled by the team under review.
  • The project is MIT licensed with roughly 959 stars at recording time, and setup takes two plugin commands.

Resources

Published September 2, 2026. Writeup generated from a favorited TikTok.