A personal library of 535 expert-persona agent skills across 20 subject domains (plus one internal metadata namespace), served over MCP and usable from both Claude and Codex. It is built on one idea: an agent gets sharper when you give it a named expert's methodology, typed inputs and outputs, and a way to cite its own evidence — then let those experts call each other.
I build these to think with. The architecture is the point; the domains are where I stress-tested it.
Live: Neural Observatory · Infrastructure map
The work I care most about is turning clinical-development and regulatory judgment into agents that show their reasoning:
- Asclepius (its own repo) — an AI agent system that reasons about a single drug program end to end: probability-of-success anchored to real trial outcomes, phase-gated rNPV, an 8-pillar clinical/regulatory scorecard, and a memo that forms a falsifiable view and names the one readout that would flip it. Every number ships its citation.
skills/biotech-venture/(42 skills) — the diligence engine behind it: clinical-development planning, endpoint and trial-design critique, probability-of-success estimation, regulatory-path reasoning, and evidence-tiered synthesis.
That's the through-line in all of it: AI earns adoption when it scaffolds an expert with structured, cited synthesis — not when it tries to replace the expertise.
- One registry is the source of truth (
data/registry.json) — every skill's metadata, dependencies, and scores. CI enforces zero drift between the skill files, the registry, and the search index. - Skills compose. Orchestrator skills dispatch sub-skills through typed Accepts / Produces contracts wired into DAG pipelines, so a high-level ask (e.g. "diligence this asset") fans out to specialists and reassembles.
- Served over MCP. A read-only MCP server exposes the library to Claude and Codex — including
search_skills,get_skill,analyze_impact, and a live health/telemetry loop. Seemcp-server/and the Codex gateway guide. - It maintains itself. A QA loop tracks usage, gaps, and feedback, and opens PRs to keep the registry, wiring, and scores calibrated.
I use the same skill-and-orchestration pattern everywhere I'm curious. The biotech and research domains are the serious work; the rest are where I pressure-test whether the architecture generalizes.
| Domain | Skills | Domain | Skills | |
|---|---|---|---|---|
| investing | 56 | worldbuilding | 27 | |
| collector | 54 | binding-vow (prompt craft) | 27 | |
| world-history | 53 | game-theory | 22 | |
| professional-development | 42 | design | 17 | |
| biotech-venture | 42 | data-science | 17 | |
| sommelier | 36 | neocortex | 13 | |
| writing | 34 | infrastructure | 12 | |
| product | 31 | research | 11 | |
| philosophy | 29 | consumer-research | 7 | |
| (+ artifacts, narrative, internal meta) | 5 |
Each skill is a SKILL.md with YAML frontmatter (name, description, dependencies, evidence conventions) plus reference docs and, for many, an eval harness. Start in skills/biotech-venture/ or skills/research/; data/registry.json is the machine-readable index of everything.
Released under the MIT License.
Built independently, for my own use. Not a product.