Compound Engineering: 33 Claude Code Skills Explained

What is the Compound Engineering plugin

Compound Engineering, indexed from everyinc/compound-engineering-plugin, packs 61 components for Claude Code: 33 skills, 27 agents, and 1 command, filed under development on Flowy. Two skills spell out the name: ce-compound writes down a durable lesson right after solving something, and ce-compound-refresh later checks whether the codebase has moved past an already-captured lesson. The rest cover framing, planning, building, reviewing, and shipping work, plus sweeping feedback back in.

Flowy indexes this plugin from its public repository. It does not host, modify, or bundle the code. The install command lives on the listing page linked at the top of this guide.

How the workflow fits together

Every skill but one carries a ce- prefix, tracing a single loop, not 33 separate tools.

Framing starts with ce-strategy for a standing strategy document, ce-brainstorm to turn a vague idea into a requirements-only plan, and ce-ideate to generate and grade directions before choosing one. ce-pov gives a decisive, project-grounded verdict, backed by the pov-peer agent forming that view from evidence in the working tree. ce-prototype is the fallback: a throwaway build for when a cheap sketch cannot settle how something should feel.

With a plan set, ce-plan structures multi-step work, ce-doc-review checks a planning document with role-specific reading angles, and ce-worktree sets up an isolated worktree, new or attached to an existing branch, PR, or commit. ce-work executes a plan end to end with local verification only; lfg runs the whole pipeline hands-off through an opened PR with no check-ins.

Review follows: ce-code-review before a PR exists, ce-debug for a diagnosis loop on failing behavior, ce-simplify-code for settled code that works but reads poorly, and ce-resolve-pr-feedback for feedback already left on one, paired with the pr-comment-resolver agent, which implements one accepted fix at a time. ce-polish and ce-dogfood both drive a real browser, one on live feedback for a feature, the other a hands-off pass that fixes small breakages on its own. ce-test-browser and ce-test-xcode cover browser and iOS simulator testing.

Shipping closes with ce-commit, ce-commit-push-pr, and ce-babysit-pr to watch one PR until it merges. ce-handoff and ce-explain handle continuity: a session handoff for another agent, or a visual teaching artifact built from evidence work-recap-scout gathers by extracting rather than interpreting.

What handles feedback and metrics instead of code

ce-sweep pulls new items from Slack and GitHub Issues, acknowledges each at the source, and emits a plan lfg can run directly; when an item carries media, the media-analyzer agent turns it into one bug-report-shaped finding in the pass. ce-riffrec-feedback-analysis does the same for a Riffrec capture or other recorded screen, voice, or notes artifact. ce-product-pulse builds time-windowed reports from configured signals, ce-optimize runs metric-driven experiment loops, and ce-promote drafts launch copy once something ships. ce-proof publishes and edits markdown in a tool called Proof. ce-setup checks the plugin's own health and repo-local configuration, and ce-retune maintains the skill corpus itself: mining a baseline from the run archive, then cutting it down in measured passes against a pre-registered bar for a new model.

What the 27 agents do if you never call them by name

The skills above call them as needed. Eleven are research agents that each note the current year is 2026, so staleness judgments do not default to a training cutoff: slack-researcher, framework-docs-researcher, git-history-analyzer, issue-intelligence-analyst, and repo-research-analyst, plus six more covering sessions, web sources, best practices, stale decisions, recent files, and outside pricing or capability claims.

A second cluster reviews code and systems from an expert stance: architecture-strategist checks changes against established patterns, security-sentinel reasons like an attacker, performance-oracle looks for algorithmic and database bottlenecks, data-integrity-guardian and data-migration-reviewer cover migration and schema-change safety, and figma-design-sync checks an implementation against its Figma source. A third group supports process: spec-flow-analyzer reads a spec for missing flows and edge cases before implementation starts, learnings-researcher surfaces past decisions before new work begins, and implementation-worker implements exactly one supplied unit of work inside a bounded authority.

What the one command does

triage-prs is a slash command, not a skill: it triages every open PR with parallel agents, then labels, groups, and reviews them one by one. It is the batch counterpart to watching a single PR with ce-babysit-pr.

Who actually reaches for this plugin

Sixty-one components under one naming convention is a lot to install for a single task like writing commit messages. It reads as built for a team running enough Claude Code work that the full loop, from a vague idea through a shipped PR to a written-down lesson, is worth codifying once, including the parts most plugins skip: sweeping Slack and GitHub Issues for new work, and folding what was learned back into the next ce-plan or ce-brainstorm pass.