research-spike

Research Spike

Answer a decision-relevant question with traceable findings, a recommendation, and concrete next steps. Investigate directly; do not merely package questions for another research tool. Work without a feature catalog, companion skill, connector, specific model, or ticket system.

Frame the decision

Inspect the request, previous decisions, applicable repository instructions, relevant code, tests, dependency manifests and lockfiles, and existing research. Discover actual versions and constraints; do not assume a stack or treat a manifest's version range as the installed version. Use existing evidence before asking for information the repository can answer.

State the question, decision it enables, relevant constraints, and what evidence would settle it. Ask up to three related follow-up questions when missing information materially changes the options, acceptance threshold, experiment, or scope. Offer concrete choices and a recommendation when useful. No minimum question count or mandatory interview rounds. Preserve answered questions and authorization. For a broad request, propose a useful bounded question and investigate its unambiguous parts.

Honor explicit time, cost, or tool limits. Do not require the user to invent an hours-or-days estimate. Without a supplied budget, use a bounded evidence plan: inspect the relevant implementation, consult sources for remaining questions, and run the smallest useful experiments. Stop when decisive criteria are resolved, the budget is reached, or additional work cannot resolve the remaining gap with available access. Report uncertainty instead of extending research indefinitely or forcing a positive answer.

Gather evidence that changes the decision

Separate verified repository facts, external claims, experimental observations, inference, and unresolved assumptions. Reuse applicable evidence, checking its version, date, environment, and scope. A stale prototype or a journal's success claim is not current proof.