Research uncertainties before writing code

Auto-Research defines a consent-based workflow for resolving uncertainty during implementation work. It supports public web research and an optional ChatGPT consultation, then presents the findings as concise options with sources. Implementation begins only after the user chooses an option or approves proceeding.

The first step establishes a research boundary: the source, exact query or redacted prompt, any workspace text that would leave the machine, and likely cost. Research waits for approval of that specific boundary. The next step uses the approved search method or selected browser consultation session, followed by a summary of approaches for the user's decision.

This separates permission to research from permission to implement. A shorthand such as a question mark or double question mark does not authorize an external submission, and a prior consultation does not automatically authorize another. Each submission needs fresh approval of the exact text.

Use it when a question has several valid approaches, algorithm or API details are uncertain, architecture choices need comparison, or the user requests external research. Its examples concern optimizer design and convergence criteria, illustrating technical decision support rather than a dedicated optimization library.

ChatGPT consultation is optional. Ordinary web search is sufficient when it meets the need. Browser-based consultation requires a user-configured, pinned automation connector and separate authorization for the profile and connector version. The skill forbids automatically installing browser packages or using a moving latest version.

It does not replace domain expertise, and research quality depends on available sources. If sensitive text cannot be safely redacted, it is not submitted. When automation is unavailable, the workflow uses an approved search alternative or asks for another approved method.

Inputs include the uncertain question, candidate approaches, approved research source, exact redacted text, and any separately authorized browser setup. Output is a concise comparison with source attribution, followed by an implementation decision. A web-search request can begin with:

User: ?? ADMM convergence criteria best practices

The workflow responds by proposing a specific query and explaining that workspace files and conversation history will not be included. Only after approval does it retrieve public sources and present a criterion for review. This example makes the query boundary visible before information is sent externally.

Redaction removes secrets, personal and customer data, proprietary code, and internal URLs. Conversation history is excluded by default. A consultation tab is isolated from other tabs and profiles; cookies, saved passwords, and unrelated session data are not accessed, exported, or reused. If login is required, the user handles it.

The skill is intended for Claude Code workflows, with web search and browser automation as research paths. Playwright is named for browser interactions, and page.fill is preferred over simulated keyboard typing for inserting the approved prompt. Related skills cover systematic debugging of browser interactions and condition-based waiting for responses.

The selected connector and profile define the browser boundary. The skill preserves their state and does not alter the user's profile as part of consultation.

It suits people making implementation decisions where uncertainty justifies external evidence. It is useful when the exact information leaving the workspace and the final implementation choice must remain under the user's control.