How Usual turns old decisions into useful context.
Usual is a local workflow for the choices that sit between “the agent asked” and “the code shipped.”
Find the question and answer
Usual looks through selected local Claude and Codex conversations for moments where you were asked to choose. It preserves the question, alternatives, answer, reasons, and surrounding context.
Retrieve relevant evidence
When a similar decision comes up, Usual searches the active evidence for examples that fit the current situation. A source and line reference travel with each example.
Choose how close to stay
Autopilot makes routine reversible choices and reports them afterward. Check-in recommends a call and asks before proceeding. Escalation pauses when evidence is missing, weak, or conflicting.
Review the result
You can accept a decision, correct it, or leave it out. Only explicit acceptance or correction turns a new prediction into project-scoped evidence.
Evidence is not a personality test
Usual does not turn every “yes” into a universal rule. It keeps exceptions. It separates observed human answers from agent predictions. It keeps the original prediction when you correct it, so the audit trail remains honest.
Past approval is not current permission. Every mode still requires current authority for deletion, spending, publication, sharing, and credentials.
What Usual does not do
- It does not train a model or launch a second agent.
- It does not modify original transcript files.
- It does not promise perfect prediction accuracy.
- It does not replace the coding client's permission controls.
For a concrete walkthrough, read the reading-list example. For the implementation details, start with the installation guide or the source repository.