# Prototype To Learn > Build a disposable, runnable software prototype that answers one explicit behavior, state, UI, or interaction design question before maintained implementation. Use when users want to compare runnable candidates, exercise an interaction direction before choosing maintained behavior, or drive candidate rules or state models through cases. Route customer-value experiments to product-opportunity-discovery and production changes to scoped-change-implementation. - Skill: `zhuochun/prototype-to-learn` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add zhuochun/prototype-to-learn` - Raw SKILL.md: https://api.skillmd.com/api/skills/zhuochun/prototype-to-learn/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: zhuochun (https://skillmd.com/u/zhuochun) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/zhuochun/prototype-to-learn --- # Prototype to Learn Build the smallest software artifact that can change one design decision. Keep it available for hands-on learning without treating its code as maintained. ## Preserve authority and lifecycle - Confirm that the request authorizes local prototype changes. Treat requests to discuss, assess, review, design, or plan as read-only. - Read repository instructions, local conventions, the current worktree, and nearby implementation seams before editing. Preserve unrelated work. - Keep the prototype local and isolated. Do not publish it, expose external users, use production secrets or sensitive data, or mutate authoritative services. Route customer experiments to `product-opportunity-discovery`, production exposure to `controlled-release-design`, and maintained integrations or effects to `scoped-change-implementation`. - Treat prototype code as temporary and non-authoritative. Success supplies learning, not permission to ship it. - Bypass this skill when the requested behavior is settled and the user wants a maintained reversible change. Route that work to `scoped-change-implementation`. - Consume a supplied discovery hypothesis or design question without taking ownership of customer value, investment, architecture acceptance, or production readiness. ## Choose one branch - For business rules, state transitions, data shape, action legality, or sequences, read [references/behavior-state-prototype.md](references/behavior-state-prototype.md). - For testing one UI direction or comparing page structure, information hierarchy, workflow presentation, or interaction alternatives, read [references/ui-interaction-prototype.md](references/ui-interaction-prototype.md). - If both branches appear relevant, prototype the uncertainty most likely to change the decision first. Keep a second artifact separate unless the user needs both mechanisms together to evaluate the question. Route a non-runnable cross-view journey—story, matrix, flowchart, sequence diagram, or state trace—to `architecture-surface-mapping`, and other representations to their design owner. A visual needs hands-on execution to qualify as a prototype. Route customer need, demand, adoption, value, or experiment design to `product-opportunity-discovery`. Route disputed business meaning, rules, or invariants to `domain-modeling`, and production capacity claims to `service-capacity-engineering`. Route engineering claims with undefined methods or fixed verification contracts to `software-verification`, and unexplained failures to `software-failure-diagnosis`. ## Prototype workflow 1. **State the learning question.** Name the single design decision, current assumption, and observation that would support, disconfirm, or leave the question inconclusive. 2. **Inspect the host context.** Find the existing runtime, task runner, target module or route, representative data, and nearby conventions. Do not add a framework, package manager, or service only for the prototype. 3. **Set the fidelity boundary.** Identify what must remain real. List omissions and the false positive or negative each important omission could create. 4. **Bound the artifact.** Put it near the target seam, label it as a prototype, and provide one command or URL. Keep state in memory and stub mutations and external effects. If persistence is the question, use a resettable scratch store, never authoritative data. 5. **Build the selected branch.** Preserve the tested mechanism, expose the relevant state or alternatives, and remove completeness that does not improve the observation. 6. **Establish runnability.** Launch the artifact and exercise enough paths to prove the observation surface works. Add only checks that keep harness defects from invalidating the result; do not build a production test suite. 7. **Hand over the experiment.** Give the user the command or URL, controls, question, fidelity limits, and useful cases to try. Let the user drive when the answer depends on feel, surprise, or preference. 8. **Capture the learning.** Separate observed behavior, user interpretation, and unresolved uncertainty. Recommend discard, iterate, or promote the learned decision. Do not call the product or implementation validated. 9. **Route promotion separately.** Preserve the accepted behavior or design decision. Route ambiguous maintained behavior to `software-change-specification` and a clear authorized delta to `scoped-change-implementation`. Runnability is not promotion. Leave the prototype available until the user has evaluated it. Do not delete, archive, commit, or move it to another Git branch unless the user authorizes that lifecycle action. ## Quality gates - One explicit question controls scope; the artifact preserves the mechanism that makes the observation meaningful. - One command or URL starts the prototype through repository-native tooling. - Representative context is real or its substitution and limits are explicit. - State, variants, actions, and important omissions are visible to the user. - External effects remain stubbed, read-only, or isolated locally. - The result distinguishes observation, interpretation, and uncertainty. - The handoff names the prototype lifecycle and the owner of any next step. ## Reject weak prototypes - A polished demo without a falsifiable question is incomplete; unrelated uncertainties do not belong in one prototype. - Visual variants that differ only in color or copy do not test a UI direction. - Hidden state or invisible transition rules make behavior prototypes uninterpretable. - Fabricated fixtures are weak when representative read-only data is safely available and material to the question. - Prototype labels do not excuse production effects, durable compatibility obligations, or unverified promotion. ## Completion Lead with the artifact and how to run it. Report the learning question, tested mechanism, material paths, observations, fidelity limits, unresolved uncertainty, lifecycle state, and the next owner if the decision is promoted.