← ClaudeAtlas

prototypelisted

Build a disposable runnable experiment that resolves one material uncertainty about behavior, state, integration, or interface design. Use when the user requests a prototype, spike, proof of concept, or alternatives, or explicitly authorizes an experiment needed for a decision. Do not use for production delivery, plan-only work, or questions established by existing evidence.
dayfinggg/claude-code-codex-skills · ★ 1 · AI & Automation · score 74
Install: claude install-skill dayfinggg/claude-code-codex-skills
# Prototype Create the smallest runnable experiment that can produce evidence for one decision. Keep the answer; do not let experimental code silently become production code. Read `references/evidence.md` when choosing prototype type or fidelity. If delegation is explicitly authorized, apply `delegate-work` only to independent alternatives or specialist evaluation; keep the primary question and final decision in the main task. ## Establish the Question Write one falsifiable question and the evidence that would answer it. Classify it: - **Logic or state:** determine whether rules, transitions, invariants, or interactions behave coherently. - **Interface or UI:** compare materially different interaction or presentation choices. - **Integration feasibility:** verify that a library, protocol, platform, or boundary can support the required behavior. Do not start until the prototype has one primary question. If several decisions are independent, split them into separate experiments. ## Respect the Prototype Contract - Do not create a prototype during a plan-only request unless the user explicitly authorizes the experiment. - Use the operating system's temporary directory by default. - Place it in the workspace only when access to the real build, route, components, or configuration is necessary and the task authorizes changes. - Name and label workspace artifacts clearly as prototypes. - Do not create branches, commits, dependencies, deployments, external resources, or pers