Capture your first real-work episode
This workflow is intentionally small. The first milestone is not a benchmark dashboard—it is one task whose real outcome you can defend.
Prerequisites
- Node.js 22
- pnpm 10
- A local checkout of the Merion repository
- A real task that has not already been solved
1. Link the private-alpha CLI
From the Merion repository root:
The linked merion command works from any folder. While changing CLI source, keep this command running in the repository:
2. Initialize a factory
A factory is the durable learning boundary for one real workflow. It can begin in a missing folder, an ordinary directory, or a Git repository.
For an explicit, agent-friendly setup:
Initialization creates private .merion state. It does not initialize Git, create a GitHub remote, or publish your evidence.
3. Ask Merion what comes next
merion next returns one evidence-backed action, why it is next, and any decision that must remain human-accountable.
4. Start genuine work
Start the episode before its solution is known:
Merion asks for the original task, workflow family, recurrence, economic value, and harness. Continue doing the work in your existing harness.
For an agentic caller:
5. Close the outcome loop
After the work is complete:
The closure interview records whether the result was accepted, what required correction, what would have made it unacceptable, and how an independent reviewer could recognize success.
Success criterion
You are done when the episode contains a defensible starting state, a real trajectory, and human-reviewed outcome evidence. Continue with capturing real work to understand what makes an episode useful for evaluation engineering.Last modified on August 21, 2026
