prototype-feature

To generate competing prototypes from an OpenSpec change, the skill dispatches N parallel variant agents using isolated git worktrees. It creates multiple working feature skeletons steered by distinct angle prompts to prevent cross-contamination.

Parallel prototyping applies cheap multi-phase validation scoring by running smoke and spec tests against each skeleton. Validation outcomes are recorded as data rather than hard-fail exceptions, allowing objective comparison between variants.

Yes, you can converge prototype variants through human pick-and-choose feedback by selecting per-aspect winners like data model, API, and tests. The skill synthesizes your selections into a best-combination convergence plan.

The best way to structure prototype findings for iteration handoff is writing a single prototype-findings.md file. This file persists human picks alongside VariantDescriptor-compatible JSON for downstream convergence processing.

Yes, you need git worktrees to run parallel prototype divergence because they provide isolated branches for each variant agent. This isolation prevents cross-contamination between competing feature skeletons during generation.

A limitation of automated variant synthesis is that it requires strict input validation for variant and angle counts. Additionally, vendor-diversity assignment is only best-effort, and full implementations are deferred until human convergence is complete.