| name | Compound Engineering |
|---|---|
| last_updated | 2026-06-04 |
Engineering work with raw agentic tools doesn't compound — each task restarts from zero, hard-won planning and lessons aren't captured or reused, and new Claude Code capabilities go underexploited because there's no system that turns yesterday's work into today's leverage.
The compounding loop is the bet: every skill writes durable artifacts (STRATEGY.md, plans, learnings via ce-compound) that the next skill reads as grounding, so leverage accrues in the repo instead of evaporating per-session. The plugin stays on Claude Code's bleeding edge so that loop always runs on the strongest available primitives.
Primary: The single developer fluent in agentic tooling who treats Claude Code as their primary engineering surface. They're hiring Compound Engineering to turn a pile of one-off agent invocations into a system — capturing plans, reviews, and learnings so each task compounds — and to ride new Claude Code capabilities without rebuilding their workflow each time.
- Loop adoption — share of real tasks that run the chain (plan → review → compound) vs. one-off skill calls; measured from session history.
- Learnings reuse — how often a prior
ce-compoundlearning is surfaced and applied in a later task; measured from repo artifacts and session history. - Rework / churn rate — how often plan, review, or work output needs redoing; qualitative today, not yet instrumented.
Adopt new Claude Code primitives fast — dynamic workflows is the live example.
Why it serves the approach: Keeps the compounding loop running on the strongest available primitives.
Deepen the artifact chain (strategy → plan → review → compound → pulse) so leverage accrues across tasks.
Why it serves the approach: This is the core bet — it's the mechanism by which work compounds.
Keep the converter shipping the plugin to Codex, Cursor, Copilot, and other targets.
Why it serves the approach: Lets the loop run wherever the developer works, not just in Claude Code.