Explore

Use the work-artifact contract throughout; keep the owning issue current as understanding changes. Reuse current guidance already in context, but read the latest artifact at the start.

Establish the current problem from the work artifact, local Markdown, prototype, or code. Name the uncertainty, the working belief to test, and the choice it could change. Identify plausible observations that would strengthen, narrow, or overturn that belief, including what would leave it unresolved. If evidence and conventions already support a direction, reconcile that judgment without manufacturing alternatives or a prototype. Otherwise choose the smallest useful trial: a few meaningfully different options or one focused discriminating test.

Apply the Development Principles: build lightweight, reversible working prototypes in the actual product when feasible. Run them and exercise the uncertain behavior; for UI work, use the browser to experience the flow, state changes, and relevant failure cases. Keep enough trial conditions and observations in working notes to interpret the result, including shortcuts that could explain it. Writing and running code can expose constraints the proposal missed; reconcile discoveries before choosing the direction. Screenshots and static mockups can supplement a working prototype but cannot establish its behavior.

Make the prototype available for the person to test: provide a working preview or runnable entry point, the short interaction and result to observe, and material shortcuts or limitations. Invite their observations when experience or preference could determine the choice; leave that choice open when their judgment is needed. Do not claim human testing occurred because the agent exercised the prototype. For work without a UI, provide the equivalent runnable operation and observable result. If a running prototype is impractical, state the constraint and use the closest useful evidence. Name shortcuts that need refinement before production.

Compare the observations with the working belief and explain what they support changing or retaining. Successful execution alone cannot justify a direction or establish user value. Choose only as far as the evidence supports; for mixed or inconclusive results, preserve the uncertainty and name the next observation or judgment that could change the choice. Use supported learning to refine scope and dependencies, reading the delivery-slicing guidance when broader work needs decomposition. Preserve uncertainties that affect boundaries without fixing speculative follow-ups.

Primarily strengthen Solution + Decisions + Reality within their existing limits: Solution holds the supported approach, Decisions connects consequential choices or removals to what was learned, and Reality holds the observations and their limits. Distinguish observations from interpretation, hypotheses, planned tests, or planned feedback; test accounts are not customer feedback. Retain only alternatives whose reasoning still affects judgment, with raw trials in working notes rather than a new experiment section.