Search

Complete the requested inquiry with evidence the caller can actually use. A result may be a precise answer, a comparative judgment, a mechanism, an experience reference, or a changed question. Understanding can be the whole deliverable; it need not become code or a report.

Recognize when looking outward matters

Recover the requested outcome, settled choices, live proposals, relevant evidence, and authority already granted. An explicit request to research is one trigger, not a prerequisite. Distinguish a small inquiry inside another task from a complete research commission. Local files and prior decisions provide context; inspecting them is not itself an external inquiry.

Search when at least one of these conditions is present:

• A substantial product, form, interaction, visual, architecture, or implementation choice remains open. Inspect relevant implemented systems, observable works, mechanisms, or adoption conditions before committing; prior evidence that one would help is not required.
• An unresolved external fact, current capability, policy, provider behavior, or implementation detail would materially change a local decision, even when the local task is small.
• The user explicitly asks to research, compare, verify, find references, or explore what else is possible.
• An open creative direction lacks concrete reference points, or the current candidates repeat the same familiar pattern.
• Repeated rejection or changes at one layer leave the cause unclear, suggesting that the problem frame, interaction relationship, execution, or candidate fidelity needs to be distinguished.
• A new observation, experiment, or feedback opens a relevant direction that did not previously exist.

Neither a clear goal nor confidence in a familiar solution settles the solution space. Relevant prior references can be reused when their conditions still apply, and discovery can reopen during implementation when a current choice stops being settled. Do not search merely because a project is new, no reference was supplied, or research would make the work look more thorough. Work proceeds directly only when its scope is settled and its applicable evidence is already sufficient; that exception covers a clear small repair and a faithful implementation of a settled choice.

When the materials are already present and the job is to explain their relationships, Study is the relevant capability. When the job is to develop a more interesting possibility, Explore may own the commission; when an interpretation or choice needs to become judgeable, Shape may help. Use the needed method without forcing a transfer ceremony or restarting the task.

Read proactive inquiry and reference choice when deciding whether or how to inspect a substantial open choice, when comparing multiple references, or after repeated rejection. For complete constructed walkthroughs of multiple-reference judgment or repeatedly rejected work, read proactive inquiry examples.

Put effort where the question needs it

Infer the intent from the requested outcome:

• Verify: settle a specific factual or capability claim.
• Compare: cover the differences that could plausibly change a choice in this context.
• Discover: expand the possibility space with concrete, relevant references or mechanisms.

These intents may change as evidence arrives. A comparison may expose a missing criterion; an open discovery may become a precise verification question. Do not preserve a stale research frame for procedural consistency.

Search breadth, reading depth, synthesis effort, response size, and retention are different choices. Start a clearly complex comparison at the depth it needs. A narrow fact may require deep tracing; a creative reference may require close observation of only a few works. Keep the user-facing result as concise or extensive as the commission warrants.

Fit the evidence to what must be judged: direction through observable works, experience through the actual medium or interaction, mechanism through source, dataflow, tests, or a probe, and adoption through local interface, state, maintenance, version, and license conditions. Keep source claims, direct observations, inference, and local transfer distinct.

Choose available native search, connectors, browsers, or established providers by the capability needed. Retrieval, known-URL extraction, corpus lookup, and model-generated research are different operations. A provider switch must not silently turn original material into generated synthesis or ignore required language, date, or domain filters. Do not install tools, create credentials, or add paid routes merely because a reference mentions them.

Read source capabilities when the inquiry spans corpus discovery, sustained research execution, or visual/interactive references. Let a concrete commission justify useful new capability without turning the current toolset into a permanent ceiling.

Read evidence-and-stopping when selecting sources, deciding what to do next with the evidence, or evaluating whether the inquiry is complete. Read returning-to-the-work when findings from a handoff, earlier task, or external inquiry must influence an ongoing project.

For long sources, repeated reading, or cross-session evidence reuse, read evidence-reading. It explains how to keep conditions visible and use the optional local snapshot helper. Ordinary search does not require it.

For a research handoff across sessions or harnesses, read research handoffs. Its optional local packet helper carries an explicit brief and selected snapshots without depending on the original store.

Preserve what the source actually contributes

Separate:

• what the source or observed behavior establishes;
• what you infer from it;
• why that matters here.

Preserve the conditions that make a claim true: version, plan, population, date, units, exclusions, and observed context. A verbatim excerpt can still mislead when its qualification has been omitted. Search rank, similarity, and agreement among retrieval providers describe retrieval, not independent corroboration or factual confidence.

Do not reduce every source to an adoption decision. A project may contribute one useful interaction while its architecture remains unsuitable. A work may change the quality being pursued without supplying reusable code. A negative result may close a route or strengthen the existing direction.

For an experience-facing choice, inspect the consequence a person will actually encounter when feasible: the delivered email, exported artifact, waiting page, interaction, or rendered output. Feature tables and developer-facing setup notes may omit the deciding difference.

Return to the original work

When the inquiry's requested outcome is supported, return to the original scene and make the effect explicit:

• A verified fact corrects the basis of the current decision.
• A mechanism is adopted, adapted, composed, or probed under the conditions that make it work.
• An experience reference remains visible while its valuable relationship is tested in the real deliverable.
• A reframing explains which earlier assumption changed and which choices are now genuinely open.
• Insufficient or unsuitable evidence leaves a bounded unknown; do not manufacture novelty to justify the search.

Carry the learned relationship into the affected work rather than extracting a feature name. When returned findings must change the project, follow returning-to-the-work rather than producing a detached summary. If the commission is research-only, complete that research without turning it into implementation.

Stopping one query does not complete a comparison or research commission. Check the requested dimensions, material conflicts, and useful counterevidence before claiming completion. Conversely, do not keep searching after the answer is adequately supported merely to fill a source count. If authorized exploration is still producing materially different possibilities, continue developing them instead of stopping at the first actionable idea.

Continue within the user's existing authorization. A local probe is appropriate when it is permitted and observation would answer the question better than more reading. Read-only research remains read-only; the Skill name grants no implementation, installation, purchase, account, production, or publication authority.

Ask the user only when a newly exposed choice belongs to their authorship or values and has not already been delegated. Otherwise make the supported reversible decision and continue. Keep any answer scoped to the next action it actually supports.

Yield to the active owner

When Search owns a research commission, deliver its complete result. When a specialized method owns a larger task, return the evidence and its implications to that method without producing a competing brief. Domain-specific evidence standards remain authoritative. Use shared working context when methods need to continue the same line of work.