Skills reference
The loop ships as Claude Code skills, installed project-scoped to .claude/skills/ when you run systemix init. Run them with /skill-name.
14 Claude Code skills ship with the loop. This table is generated from the skills themselves at build time — it can't drift.
| Skill | What it does |
|---|---|
/atlas | Interview the instance's topology (ICP, experiment surfaces, signal, cadence), then generate a topology-aware multi-agent loop workflow (.claude/workflows/<slug>-loop.js) plus a thin runner skill. The workflow wraps `systemix loop` (propose-only, never closes) and adds ICP-lens surface audits. Re-run with `update` to regenerate after the topology changes. |
/build-workflow | Turn a short guided conversation into a visual, composable agentic workflow — a steps+edges graph saved to .systemix/workflows/<slug>.json that renders via the Atlas StepNode graph (and can be pasted into an experiment's `workflow` frontmatter). Covers BOTH use cases the graph vocabulary serves — an internal workflow around Systemix itself (scope:"internal"), or a persona/JTBD-driven workflow for the product being built, authored as part of the build chapter (scope:"product") — and tags every workflow with which one it is. You pick the trigger, the step order, which steps are human-in-the-loop, and where it branches or fans out; the skill composes valid steps from the fixed StepKind vocabulary and wires the edges. Propose-only — the workflow file is written only after you approve it. |
/close-experiment | Close a running experiment — record the result + decision in experiments/<id>.mdx, append the learning to experiments/LEARNINGS.md (the Memory ledger), and queue a HITL decision card in .systemix/queue.json. File-first (engine = Claude Code); no external service required. Use when PostHog evidence is decision-ready. |
/connect-signal | Connect a data source (PostHog today) so experiments can gather live evidence. Guides the wiring, verifies it with `systemix evidence check`, then flips the signal on in systemix.config.yaml. Run when you're ready to measure — experiments author + run without it. |
/growth-audit | Audit running experiments against PostHog evidence. Reads experiments/ for running experiments, fetches cached evidence from the experiment files, and produces a prioritized decision brief. Use weekly to close the evidence loop. |
/hermes | Trigger Hermes to synthesize evidence from an experiment and queue a HITL decision. Engine = Claude Code (no local model). |
/init-experiment | Start the next experiment end-to-end — frame the bet via tight questions, scaffold via the CLI door (systemix experiment new) and enrich the MDX, wire the metric for real, ship through the existing variant seam, verify, and open a PR. Use before writing any variant code. |
/jtbd-audit | Run a Jobs-to-be-Done + Outcome-Driven Innovation (ODI) analysis on any product — interview the core job, capture desired outcomes in canonical ODI form, score each by importance vs satisfaction, and rank the underserved opportunities. Generalizes docs/product/jobs.yaml (Systemix's own JTBD source of truth) into a reusable methodology; the opportunity ranking seeds what to build and which hypotheses to test. Read-only; any jobs file it would write is proposed for approval. |
/measure | Add PostHog instrumentation to a component. Reads the experiment contract for what to measure. |
/new-goal | Declare a new goal in experiments/goals/ — the north-star outcome a set of experiments pursues. Humans give goals; the engine never self-assigns one. Use before framing an experiment that doesn't yet have a goal to serve, or when the instance's priorities change. |
/propose-experiment | The engine's generate stage — propose the ONE next experiment from earned evidence. Reads the digest (LEARNINGS recall, running experiments, goals, ODI opportunity ranking) via `systemix propose context`, drafts a single falsifiable ICP-calibrated bet, and queues it as a hypothesis-proposal card via `systemix propose queue`. Propose-only — a human approves the card on Home (/config); approval scaffolds the contract, and nothing ships until /write-variants and a merged PR. Runs headlessly on the weekly engine cron or on demand. |
/systemix-init | Initialize a fresh Systemix instance and co-author experiment |
/workflow-artifact | Turn a saved workflow (or an experiment's `workflow` frontmatter) into a portable, shareable Atlas view — a single self-contained artifact that renders the step-node graph with no repo, dev server, or network access required. The workflow data is baked in at generation time (a snapshot), so it works anywhere. Read-only; it renders an existing workflow, it does not author or run one. |
/write-variants | Generate and implement variant copy or UI changes for a running experiment. Reads the hypothesis contract for context, proposes new variant text calibrated to the ICP, and optionally writes it to the relevant source files. Use to iterate on an experiment or add a third variant. |
Three-door equivalents
The same operations work from the CLI and MCP, without Claude Code:
# CLI
systemix experiment new
systemix experiment list
systemix experiment measure <id>
systemix experiment close <id>
systemix experiment learnings
systemix loop# MCP tools (for any agent)
experiment_new
experiment_list
experiment_get
experiment_measure
experiment_close
experiment_learningsChanging the loop itself
/hermes can propose changes to the loop's own skills and rules. Those proposals always arrive as a card asking for your OK — they are never applied automatically, even when the instance is set to run autonomously.