Compound Workflow

A portable, command-first system for shipping software with less ambiguity and stronger verification.

Inspired by Compound Engineering — the idea that small, consistent improvements compound over time into significantly better outcomes. Applied here: every piece of work follows the same cycle, every decision is captured, and every session makes the next one faster.

What it does

Compound Workflow gives your AI agent a structured process for turning a request into validated, documented output. Instead of ad-hoc conversations, you get a repeatable cycle:

Clarify → Plan → Execute → Verify → Capture

Each step has an explicit gate. Work doesn't move forward until the previous step is done. Learnings are captured so the next similar problem takes minutes, not hours.

Get Started

npm install compound-workflow

Postinstall runs automatically and sets up your repo: AGENTS.md, standard directories, and OpenCode wiring.

If your package manager skipped the lifecycle script, or after updating the package:

npx compound-workflow install

Then run /setup-agents in your agent harness to tailor AGENTS.md to this project — it detects your stack and installed harnesses, curates the Skill Index, and prompts for anything it can't infer. Re-run it whenever harnesses or skills change.

The Workflow

brainstorm → plan → work → review → compound → metrics

/workflow:brainstorm

Clarify what to build before any planning starts. The agent asks structured questions, surfaces ambiguities, and produces a brief that confirms intent and constraints. Nothing gets designed until the "what" is agreed.

/workflow:plan

Turn the brief into an executable plan. Declares fidelity level (Low / Medium / High), confidence, solution scope, and open questions. Cites local code references and prior solutions. No code is written here.

/workflow:work

Execute the approved plan. The agent starts with an Opening Sequence that asks which execution context to use, recommends a dedicated worktree, and waits before reading the full plan or mutating state. After isolation preflight passes, it derives todo contracts, applies the standards gate, runs triage, delegates to subagents, and collects verified output. Evidence is required before any todo moves to complete.

/workflow:review

Independent quality check before the work is considered done. Runs in a separate pass from the implementer. Emits a clear pass / fail with standards compliance noted. Required for code and config changes.

/workflow:compound

Capture what was learned. Writes a structured solution doc to docs/solutions/ with searchable YAML frontmatter. The first time you solve a problem takes research — document it and the next occurrence takes minutes.

/metrics and /assess

Log session outcomes and spot trends. /metrics records a single session. /assess rolls up recent entries and proposes concrete improvements to commands, skills, or config.

Optional Commands

/workflow:triage — Manually curate and prioritise a backlog before execution. /workflow:work runs triage automatically, but this gives you explicit control when the queue is complex.

/workflow:tech-review — Technical correctness check on a plan before build. Use when the approach needs a second opinion before committing to implementation.

/test-browser — Browser-level validation of affected routes. Useful for UI changes where automated tests aren't enough.

/setup-agents — Tailor AGENTS.md to this project. Detects your stack and installed harnesses, curates the Skill Index, and prompts for anything it can't infer. Re-run whenever harnesses or skills change.

/install — Re-run setup. Safe to run again after updating the package or adding a new harness.

How it's structured

Skills and agents are the internals. Commands are the public API — stable entry points that compose the right skills for each phase.

| Layer | What it is | | --- | --- | | Commands (src/commands/) | What you invoke. Orchestrate the workflow steps. | | Skills (src/skills/) | Focused capabilities loaded by commands when needed. | | Agents (src/agents/) | Specialised subagents for research, review, and analysis. | | AGENTS.md | Project config: repo settings, harness list, skill index. |

On install, files are copied into whichever harness directories are present in your project (.agents/, .cursor/, .claude/). No symlinks, no generated manifests.

Guardrails

• Independent review policy: code/config changes require /workflow:review before workflow completion. Docs-only changes are exempt.
• Standards baseline policy: /workflow:work and /workflow:review both apply the standards baseline. Violations block completion.
• No ad-hoc artifacts — output goes to docs/plans, docs/solutions, todos, or docs/metrics. Nothing else.

If docs conflict: follow docs/principles/workflow-baseline-principles.md, then AGENTS.md, then command docs, then skill docs.

Learn More

• Workflow principles
• AGENTS.md — command index, repo config, skill routing
• Commands — full command specs
• Skills — skill docs and templates