# How Winskel breaks down work

> Most objectives stay with one model. Larger or riskier ones become a small set of connected steps that Winskel runs and shows you.

- Canonical page: https://www.winskel.com/docs/decomposition
- Applies to: Winskel public download 0.7.0-alpha.1 (alpha channel), Apple Silicon, macOS 12 or later
- Content updated: 2026-10-06
- Agent reference: https://www.winskel.com/agents.md

## When work is split

Splitting is conditional. Every handoff costs context: the next model has to rebuild what the previous one already understood. So by default an objective stays one task on one strong model that investigates, changes, tests, and repairs its own work.

Winskel decides when an objective needs a plan. Today it plans the work with Claude Code, and may split it, run parts at the same time, or add a review, when:

- The work involves security, authentication, migrations, infrastructure, or architecture
- The objective is too large for one step
- A step fails and needs recovery
- You ask for separate or parallel parts, or name both Claude Code and Codex
- You name an agent system or a model for one part of the work
- You ask for a review

A plan splits the work only where parts are independent, need different strengths, or benefit from an independent review. A numbered list in your objective is not on its own a request to split; write "as separate steps" or "in parallel" when you want that.

## Dependencies

Independent parts run at the same time in separate worktrees. A later step starts on its own once the steps it depends on passed their observed checks. When a check could not be observed, Claude Code reviews the result before Winskel continues. The graph shows declared execution structure, not private model reasoning.

## A worked example

For a bug fix that moves between Claude Code and Codex, with the context passed at each step, see [How to use Claude Code and Codex together](https://www.winskel.com/docs/claude-code-codex-together).