CE System Design: From Workflow to AI Agent Engineering

1. View CE as a System, Not Just a Skill Set

From the repository we see three content layers:

plugins/compound-engineering/skills/

plugins/compound-engineering/agents/

docs/brainstorms/

, docs/plans/, docs/solutions/ The main workflow, defined in plugins/compound-engineering/README.md, is:

/ce:ideate -> /ce:brainstorm -> /ce:plan -> /ce:work -> /ce:review -> /ce:compound

This is not a simple "A then B" diagram; it is a clear system hierarchy: ce:brainstorm – defines the problem (WHAT) ce:plan – decides how to build (HOW) ce:work – schedules execution ce:review – discovers risks and converges ce:compound – consolidates knowledge for reuse

2. First‑Level Insight: Complex Tasks Cannot Be Handled by a Single Large Prompt

2.1 ce:brainstorm – Define WHAT, Not HOW

The skill file plugins/compound-engineering/skills/ce-brainstorm/SKILL.md states that brainstorming answers “WHAT to build”, while ce:plan answers “HOW to build”. The output is a durable requirements document. It emphasizes right‑sizing ceremonies, keeping product decisions in the brainstorm, and avoiding implementation details in the requirements.

2.2 ce:plan – Produce a Decision Artifact, Not an Execution Script

Unlike Superpowers, which forces a fully detailed script, CE’s plugins/compound-engineering/skills/ce-plan/SKILL.md says a plan should capture approach, boundaries, files, dependencies, risks, and test scenarios, but must not pre‑write code or shell commands.

2.3 ce:work – Execution Control, Not a Coder Prompt

The skill file plugins/compound-engineering/skills/ce-work/SKILL.md treats execution as a set of sub‑problems: input type, complexity level, need for branch/worktree, task‑list creation, inline vs. serial vs. parallel sub‑agents, and explicit test coverage. It reads the plan’s Implementation Units, Requirements Trace, Test Scenarios, and Verification, and performs a system‑wide test check.

3. Second‑Level Insight: Risk Discovery Requires a Structured Review Pipeline

The review skill plugins/compound-engineering/skills/ce-review/SKILL.md defines a multi‑stage pipeline:

Define scope (PR, branch, base ref)

Select mode (interactive, autofix, report‑only, headless)

Activate specific reviewers

Collect structured findings

Orchestrate merge, deduplication, and routing

It separates “severity” (how serious) from “routing” (who fixes, can it be auto‑fixed), avoiding the common pitfall where a single reviewer both discovers problems and decides remediation.

4. Third‑Level Insight: Documentation Is a Core System Component

Durable artifacts live in the docs/ folder: docs/brainstorms/ – requirements docs/plans/ – technical plans docs/solutions/ – reusable learnings

Example files such as

docs/brainstorms/2026-03-29-testing-addressed-gate-requirements.md

and

docs/plans/2026-03-29-001-feat-testing-addressed-gate-plan.md

illustrate how requirements define problem, boundaries, and success criteria, while plans trace those requirements to implementation units and verification steps.

5. Design Takeaways for Building Your Own Agent System

5.1 Design the Backbone Before Adding Specialists

Define the core chain: problem definition, decision formation, execution scheduling, risk discovery, and knowledge consolidation. Without a backbone, adding many specialist agents leads to chaos.

5.2 Assign a Single Responsibility to Each Skill

For example, ce:brainstorm only defines problems, ce:plan only decides on approaches, ce:work only schedules execution, and ce:review only evaluates outcomes.

5.3 Ensure Intermediate Artifacts Are Reusable

Requirements, technical plans, review findings, and solutions must be consumable by subsequent steps and future iterations.

5.4 Prioritize System Design Over Prompt Tuning

CE focuses on role separation, context flow, result routing, and knowledge retention rather than crafting a single “magic prompt”. This systematic approach yields more maintainable and scalable agent pipelines.

Conclusion

Viewing CE as an AI‑agent engineering system reveals a disciplined method for decomposing real engineering problems across layers, roles, and artifacts, turning ad‑hoc prompts into a sustainable, closed‑loop system.

References

https://github.com/EveryInc/compound-engineering-plugin