MojaCode

A personal studio
for AI coding.

On the deskIndex

08 / Model switching

A new model. A continuous practice.

Changing models should not mean reconstructing the project from memory. Prepare a brief that another assistant could read without inheriting every assumption from the previous conversation.

Capture the current state

Record the objective, relevant files, existing changes, and checks already performed. Include one unresolved question and the next permitted action. Ask the new assistant to inspect the actual project before accepting the handoff as fact. The working tree may have changed since the note was written.

Compare like with like

For a comparison, use the same bounded task, input, and acceptance criteria. Record the model identifier and configuration you can verify. Review correctness, scope discipline, and the usefulness of uncertainty. A single successful example is evidence about that task, not a universal ranking of models or a reliable performance benchmark.

Keep the accepted work separate

Preserve changes you have already reviewed and identify experimental output clearly. Do not let a fresh model replace working code simply because it prefers another style. If results disagree, use the project's behavior, tests, and documented contracts to resolve the difference. Return to one concrete question rather than asking models to vote on each other.

A model handoff

Current objective: [observable result].
Reviewed changes: [files and behavior].
Checks completed: [command and result].
Open question: [one uncertainty].
Next action: inspect the current files and propose a focused check; preserve existing work.
Continue the practiceThe studio