Guides

OmO gives you several entry points. Most days you only need one of them. This page is a quick map of when to reach for each.

Choosing a Workflow

Small fix or direct question. Lowest overhead — just talk to Sisyphus.

Complex but you want autonomy. Sisyphus explores, delegates, implements, verifies, and continues.

Complex and precision matters. Plan first with @plan, then /start-work hands the plan to Atlas.

Deep GPT-native reasoning. Autonomous worker for architecture, debugging, and synthesis.

One narrow implementation unit. Spawn a focused Sisyphus-Junior with category and skills.

ulw is the "just do it" path. Give the goal and, when possible, the surface that proves it works.

ulw fix the login error message and verify success/failure in the browser

A good ulw prompt names three things:

Use Prometheus when correctness matters more than speed.

Stay in Sisyphus and type @plan, or switch to Prometheus with Tab.

@plan refactor the auth module without behavior changes

Answer the interview questions until requirements are clear. Prometheus writes a plan under .sisyphus/plans/.

/start-work

Atlas reads the plan, creates boulder state, delegates execution units, and can resume later if the session ends.

If /start-work reports no active plan, create a Prometheus plan first or clear stale .sisyphus/boulder.json state.

Aspect	Hephaestus	Sisyphus + `ulw`
Model style	GPT-native autonomous deep worker.	Orchestrator that delegates by category and skill.
Best for	Architecture, hard debugging, cross-domain reasoning.	Most "please handle it" complex tasks.
Planning	Self-plans while executing.	Uses Prometheus plans or autonomous exploration.
Delegation	Leans on Explore and Librarian for context.	Uses category-based workers and specialists.

Default to ulw. Switch to Hephaestus when you specifically want GPT-5.4-style deep reasoning.

The upstream README recommends letting an agent install OmO. When you delegate the install, the agent should:

The agent should never silently change model settings or disable hooks/MCPs.

Sisyphus runs best on Claude Opus, Kimi K2.5, or GLM 5 style models.
Hephaestus, Oracle, and Momus are GPT-friendly or GPT-native.
Visual engineering benefits from Gemini-style frontend reasoning.
Explore and Librarian should stay fast and cheap.
fallback_models and runtime fallback move past failures instead of stopping the workflow.