Prototype the unknowns, not the whole roadmap
Use a short-lived working artifact to answer named questions about an interaction, data flow, or state model; retain the decision record even if the code is discarded.
Know Today
👀 Keep an eye on
- Executable prototypes are becoming a practical planning artifact for ambiguous product behavior, not merely a late-stage demo.
- Generating deliberately different UI or interaction variants exposes choices such as filtering, grouping, and reset behavior that prose specifications hide.
- Small terminal harnesses can make difficult state transitions testable before a service or workflow is productionized.
- Keeping a prototype on a disposable branch gives later implementation work a concrete reference while preserving a clean path to discard it.
- High-fidelity agent work has a real context and token cost, so it is best reserved for questions that discussion cannot settle reliably.
🧭 What changed
Compare behavior instead of debating descriptions
Ask for a few meaningfully different approaches, run them in the relevant application context, then choose based on observed behavior and constraints.
Separate discovery from implementation
One work item can establish scope and assumptions, while another produces the prototype that tests them; this keeps uncertainty visible instead of burying it in a large ticket.
🛠️ Practical workflows
- For a complex feature, write down three to five unanswered questions before asking an LLM to build anything.
- Run prototypes where their real dependencies and surrounding UI can reveal integration friction, but isolate data, access, and rollout paths first.
- Use a CLI simulator for domain logic when the important uncertainty is transition order, retries, invalid states, or recovery behavior.
- When continuing a long agent session, preserve decisions and test results in compact notes rather than relying on the full conversation history.
💭 Opinions worth testing
Opinion: Spec-first is often overused
A working prototype may resolve interaction uncertainty faster than ever-more-detailed prompts or requirements. Test this on work where feedback depends on seeing the behavior, not just reading it.
Opinion: Prototype-to-production can be an easier handoff
A prototype carries decisions and sometimes reusable code, but its value depends on architecture, tests, security requirements, and disciplined cleanup; measure rework avoided before making it policy.
Prediction: Visual exploration may regain importance
Better agent connections to design and canvas tools could make wireframes more useful again. Treat this as a direction to monitor, not as a broadly proven capability.
⚠️ Caveats
- Reported gains are practitioner experience, not comparative evidence that the workflow will fit every codebase.
- Prototype code can conceal weak tests, unsafe assumptions, and production-readiness gaps even when it looks convincing.
- Trying changes on a live route requires prepared isolation, access controls, rollback, review, and patience; do not expose real users or data casually.
✨ Try this today
Build three versions of one awkward interaction
Choose a small feature with unclear filtering, empty-state, or selection behavior and ask an LLM for three intentionally distinct implementations. Run each locally and record which behavior answers the real user question best.
Make a state-machine playground
Describe a troublesome workflow to an LLM and have it create a small local CLI harness with valid and invalid transitions. Drive edge cases manually and turn the surprising cases into production tests.
Use worktrees for parallel prototype review
Create separate Git worktrees for a proposed implementation, a competing variant, and a review branch, then run local instances on separate ports. This keeps agent-assisted changes from sharing a working directory while you compare behavior.
Time-box a prototype handoff
Give an LLM a narrowly scoped prototype task and a disposable branch, then ask it to leave a short decision log and cleanup list. Compare the time to reach a reviewable implementation with your normal spec-first approach.