Pipelines and Hooks

Deterministic workflows that execute without model negotiation. The formatter runs before code hits disk. The linter runs after. No arguing required.

The Problem with Instructions

“Always run the formatter” is a suggestion. The model may skip it, or require reminders.

Instructions are probabilistic. Pipelines are deterministic.

What Pipelines Look Like

A pipeline chains steps where each output feeds the next input. The Unix pipe is the simplest example—cat error.log | claude -p "summarize" sends file contents to Claude and outputs the response. One command’s output becomes another’s input.

AI workflows work the same way: code written → formatter runs → linter checks → tests run. Each step triggers automatically.

Implementation Options

Tool-level hooks: Claude Code and Cursor support hooks that fire on events like “file saved” or “tool completed.”

Agent orchestration: A parent agent spawns child agents, reviews output, and routes work. The parent enforces the pipeline.

External automation: The pipeline lives outside the AI entirely. lint-staged runs formatters and linters on staged files during git commit—the model never gets a chance to skip it.