The problem
A general-purpose coding agent writes plausible code that ignores the conventions of the codebase it is editing. Worse, every engineer was re-teaching the same context to the same models, by hand, in every session — and each of them taught a slightly different version of it.
What I built
A set of 16 model-specialized agent definitions, packaged as drop-in agent kits: discovery, analysis, backend implementer, frontend implementer, code-reviewer, impact-analyzer, test-writer, and their E2E variants. Each definition pins the model, the tool surface, the output contract and the conventions for one job in the development lifecycle, so a task is routed to a narrow agent rather than a general one.
38K lines, sole author. Every one of the 16 definitions was individually blame-verified at 100% mine — this is the most cleanly attributable work in my portfolio.
How it works
Each kit is a self-contained bundle: an agent definition, its skills, and the MCP tool surface it is allowed to touch. The definitions are consumed by the Claude Code SDK and the Cursor CLI, so the same kit works from an in-process SDK run or a forked CLI binary — which is what let the platform swap coding-agent runtimes without rewriting agent behaviour.
The part that matters most is how the conventions were produced. Instead of writing down a house style from memory, I measured the real repositories — the patterns that were actually used, at the frequency they were actually used — and encoded those. An agent that is told what the codebase does is far more useful than one told what it should do.
Evidence
16 agent definitions · 38,000 lines · sole author · 100% blame-verified per definition.
What I owned
All of it. I wrote every definition, every skill and every convention in the kits, derived those conventions by measuring the real repositories, and ran the blame verification per file rather than in aggregate.
Stack
Claude Code SDK · Cursor CLI · Model Context Protocol (MCP) · Python · structured Markdown agent definitions