DTC growth engineer. Fractional CMO at Ironclad Pans.
I build tools that put constraints on models with write access to something that matters — an ad account, a distribution plan, a codebase, an afternoon. Each one ships with what it cannot do written down next to what it can.
moot — Five agents with opposed mandates argue every contested move before anything writes to a live Meta ad account. Ten mechanisms sit between a decision and your money, and the Python package that computes them never calls a model. Generalised from a production system that ran a real DTC account. 483 tests · MIT · Claude Code plugin.
fronts — Organic distribution planned as an imperfect-information game, after DeepMind's Code World Models paper. The model writes the simulator; search does the playing, and premature scaling is illegal rather than penalised. Zero runtime dependencies in the core. MIT.
glm-delegate — Token-heavy execution delegated to swarms of GLM agents in disposable git worktrees, while an Anthropic model scopes and reviews. The write guard is a tripwire rather than a sandbox, and the README says so before you install. 42 checks · zero runtime dependencies · MIT · Claude Code plugin.
discipline-hook — Checks what you are about to ask Claude to build against what you already decided to do. Reads your real priority list from markdown, Sheets, Excel or any MCP board, and journals every verdict so Friday can tell you what the week actually went on. Zero runtime dependencies · MIT · Claude Code plugin.



