The Problem

Teams building AI agents face a recurring problem: agent behavior is generic. You tell an agent to "act like a senior investor" and it produces a plausible imitation, not the actual reasoning patterns of someone who has made those decisions. Building a genuinely expert agent requires curating primary sources, distilling them, and encoding the reasoning into a format the agent will actually follow. That work is expensive and rarely shared.

What This Does

mimeographs is a collection of 80 ready-to-use agent skills, each cloning the thinking of a specific person. Each folder contains a SKILL.md (loaded on-demand) and an AGENTS.md (always-on), plus references/ files covering principles, frameworks, mental models, and quotes. The subjects span founders (Buffett, Jobs, Musk), philosophers (Aristotle, Nietzsche, Arendt), and AI researchers (Karpathy, Hinton, LeCun).

The content is generated by the K-Dense-AI/mimeo pipeline, which reads public sources, distills them with a frontier model, and clusters recurring ideas. The _workspace/ directories show the full pipeline artifacts: raw sources, distilled clusters, and verification outputs.

How It Is Wired

The repository is a static data collection, not an application. The only code is scripts/build_catalog.py, which generates catalog.json and the INDEX.md counts. The import graph is trivial: 1 internal module, 0 import edges, no circular dependencies.

The wiring that matters is the file layout convention. Each person's folder follows a strict pattern: SKILL.md for on-demand loading, AGENTS.md for persistent context, and references/ for the distilled knowledge. The catalog.json at the root provides programmatic access to all 80 experts with their trigger conditions.

The CI workflow (.github/workflows/catalog.yml) rebuilds the catalog on changes. That is the entire operational surface.

How To Use It

Clone and copy the relevant folder into your project:

git clone https://github.com/moses-y/mimeographs
# Copy a skill into your agent's skills directory
cp -r mimeographs/steve-jobs/ your-project/.agents/

The SKILL.md files follow the standard agent skill format. The AGENTS.md variant works when placed in a directory your agent operates in. The INDEX.md and catalog.json are the discovery surfaces — search by topic or match programmatically.

Real-World Use

Install the Buffett skill and ask your agent to evaluate an acquisition. The agent stops producing generic advice and starts asking about owner earnings, economic moats, and circle of competence. The references/frameworks.md and mental-models.md files contain the actual distilled patterns the agent will apply.

Code Health & Issues

Static analysis found 5 issues, none critical:

  • High - No lockfile for pyproject.toml - the artifact tested may differ from what ships. Commit a lockfile.
  • Medium - .github/workflows/catalog.yml declares no permissions for GITHUB_TOKEN - add least-privilege defaults.
  • Medium - No Dependabot/Renovate configured - dependency updates are manual.
  • Medium - No dependency vulnerability scan in CI - add dependency-review-action or osv-scanner.
  • Low - No timeout-minutes on workflow jobs - a wedged step runs to the platform default.

SDLC observations: no tests exist, and the single Python file is a build script, so the risk is limited. No secrets committed. License present.

The Bottom Line

This is a well-organized, content-heavy collection with a clear convention and working CI. The value is in the curation and distillation, not the code. If you want your agents to reason with specific expertise, this is a practical shortcut. The main risk is trusting the distilled content without verifying against the original sources — some _workspace/ folders show quote verification artifacts, which is a good sign, but you should spot-check before relying on it for high-stakes decisions.