Agent Pipeline
Coding agents are at their best when they have structure to work inside. Pipeline Skills is the structure I wanted: a sequence of seven distinct stages (Discovery, Architecture, Gameplan, Test Generation, Implementation, Review, QA Plan), each producing a structured artifact, with two human checkpoints — approve the architecture before tests are written, approve the gameplan before code is written. After that the agent runs.
It ships as a set of Claude Code skills (slash commands) that any project can install. Repos with conventions files and test infrastructure can adopt it in minutes. The setup skill auto-detects framework, stack, and directory structure and writes a configuration block into the project’s conventions file.
The principle behind it: AI-augmented development works best when the handoff points are deliberate — where humans approve, where the agent commits, where work flows between stages. Most “AI for coding” tools collapse those seams. The pipeline insists on them.
Inspired in part by Doug Engelbart’s framing of system design as the place where the highest-leverage AI work actually lives — see /writing for related thinking, and the Work Context Protocol for the persistent-memory layer the pipeline sits on top of.