Custom skills that orchestrate agent teams, scaffold entire projects, run autonomous deliberation, and analyze decision history. The infrastructure that makes work compound.
A multi-agent pipeline that produces visual artwork as code.
Modeled on a real atelier: a Scout researches references, a Composer writes the brief, a Renderer executes it, a Critic drives revision. Every piece is reproducible from a seed and a brief.
Not image generation — this skill writes code that draws. node-canvas stroke painting. GLSL fragment shaders. PptxGenJS geometric layouts. PIL pixel art. The output is a PNG you own, generated by code you can re-run, version, and modify.
The Great Wave and Gas Giant pieces in my gallery were both produced by this pipeline. Each piece lives under ~/art/<slug>/ with its research, brief, version log, generator code, critiques, and rendered PNGs — fully reproducible months later.
The system that builds the systems.
Initialize a new project with a complete 18-agent, 5-team build pipeline. Creates the .claude/ directory, agent configs, skill manifests, project rules, and tech stack scaffolding from templates — all driven by a short interview about the project domain.
Built because every new project needs the same setup work: agents, skills, rules, branch structure, test framework, CI. This skill collapses what used to be a half-day of yak-shaving into a 10-question interview.
A short series of AskUserQuestion prompts gathers: project name + slug, one-sentence description, tech stack (React + Vite default, Next.js, SvelteKit, Vue), styling and component library (Tailwind + shadcn/ui default), data layer (Supabase, Firebase, Prisma + Postgres, Drizzle + PGlite, REST), main entities and their relationships, the primary user view, and a 3-5 word visual direction.
The output is a complete .claude/ directory with agents tuned to the chosen stack, skills wired to the chosen libraries, project rules generated from the domain description, and a CLAUDE.md that reflects everything from the interview. Drop into a fresh git repo and start shipping — no boilerplate session, no stack debate.
Tooling that mines decision history to improve agent alignment.
A PostToolUse hook logs every AskUserQuestion answer to ~/.claude/decisions.db across all projects. This skill harvests the unprocessed entries and extracts actionable insights from the user's choice patterns.
The premise: a user's real preferences are buried in thousands of micro-choices they've already made. Surface the patterns, find the contradictions, propose memory candidates. The agent gets better at working with the human by learning from how the human actually decides.
Self-improving tooling for an LLM assistant. The agent that works with the user tomorrow is shaped by the agent that worked with the user today.
Multi-agent autonomous deliberation in a Model UN format.
Spawns a 12-delegate General Assembly of historical and contemporary thinkers — Peter Thiel, Naval Ravikant, Nassim Taleb, Angela Merkel, Charlie Munger, Ryan Holiday, and others — to deliberate on a single hard question. Delegates research independently, form organic blocs, draft competing resolutions, negotiate amendments, and vote on a binding outcome.
The format borrows from Model UN procedure — opening speeches, bloc dynamics, working papers, sponsor's choice. Every speech and vote is preserved verbatim. The Sponsor reviews the final resolution and can accept, modify, or reject it. The output is a full transcript plus a structured resolution document that captures both the recommendation and the reasoning behind it.
All user interaction is front-loaded into an expanded Chair's Brief. After the assembly is cast and the brief is approved, the entire debate runs fully autonomously — zero interruptions through opening speeches, bloc formation, paper drafting, amendments, and the final vote. The user can walk away for an hour and return to a fully reasoned recommendation.
Strategic deliberation on questions where the user wants multiple perspectives in tension before committing. The skill is most valuable when the question is genuinely hard, when the user has access to a deep bench of relevant thinkers to draft as delegates, and when the deliberation itself — the cross-examination, the bloc dynamics, the amendments — is as informative as the final vote.