Guru Research

Guru Research

Investigate the question with the host's chosen research tools and current sources. Keep project memory project-owned, inspectable, and subordinate to current evidence. Treat a current document or dataset supplied by the user as external research evidence; its presence in the project does not make the task memory-only.

Treat a matching natural-language request as sufficient invocation. Do not require the user to say research, Guru Maker, a Skill name, a record kind, or save. Resolve short or elliptical wording from the immediate task, explicitly supplied artifacts, and targeted project memory. Ask for clarification only when the unresolved ambiguity would materially change the research question, requested output, or authorization. When no referent and no supplied artifact exists, ask immediately without scanning project files, searching memory, or reading Guru Mode.

Resolve every linked reference path relative to this loaded SKILL.md. The loaded Skill directory is authoritative: when it is under <project>/.agents/skills/, ../../shared/... reaches <project>/.agents/shared/.... Never substitute a user-global .codex path, another plugin-cache path, or skills/shared/.... Read project retrieval before using memory, research framing for material research, research capability composition when specialized research support may fit the request, Guru Mode before a possible memory write, the workspace contract when project layout matters, and the Runtime contract before any gurumaker command. Read the in-task curation contract before retaining any result.

The host selects any research Skills, tools, and data connectors and decides how to combine them. Honor that selection and do not act as an agent or tool router. The host and specialized capabilities own research strategy, tool selection, domain procedures, evidence evaluation, and synthesis. This Skill contributes only relevant project-memory retrieval, stale-memory discipline, and retention boundaries. Apply its compact target-memory query alongside the host's chosen research path in an initialized project; Guru Mode off or a no-write instruction does not disable that read-only query.

Match the research and validation method to the question rather than assuming a company, fundamental, causal, technical, or quantitative workflow. Treat stale memory as context, not a current fact. Search only the project-owned hierarchy under guru-maker/; installed Pack source slots are never research context. Never filesystem-read a guru-maker/ record, even just its metadata before lookup; use Runtime search and exact read for all project memory.

For every material research task in an initialized project, run one compact target query before external research. Skip it only for a narrow fact or a clearly unrelated task. After candidate selection, apply the retrieval reference's bounded activation pass: exact-read an informing Decision, its authored reusable inputs, and only plausibly material supporting Evidence for evaluation; carry forward only observations actually reused. Expose missing, stale, or inapplicable inputs. If the single most consequential reusable role remains uncovered, run the one allowed concept-gap query and exact-read its selected result; do not fan out by kind or reformulate a no-match. This retrieval is useful context; the pre-write candidate search remains separately required for deduplication when Auto retains a result.

In an initialized workspace, resolve Guru Mode before a possible write. In auto, apply the shared durability gate after completing the research; do not wait for a separate save request. A new source- or calculation-bound observation with a period or cutoff that changes or materially qualifies a thesis, risk, assumption, or later review question must become exactly one primary Evidence record unless the user requested no write, the result is a narrow fact, or equivalent source-period-entity Evidence already exists. Otherwise retain only the minimum stable Wiki, well-bounded Lens, or repeatable Method improvement likely to matter again. Do not promote a one-off result or unreviewed inference into reusable memory. In off, write only after an explicit retention request. Read the record contract before writing. Use compact pre-write candidates and apply the kind-specific curation policy; new current observations remain append-oriented Evidence, while only stable reviewed knowledge improves reusable memory. Do not create or broaden a Lens from one observation; require reviewed assumptions, scope, limits, and a falsifier under the shared gate. Keep incidental cleanup within the shared five-existing-record budget, then run touched-kind health and one final knowledge check as specified there. Leave Decisions, Execution Briefs, deliberate training, and Pack lifecycle work to their respective public Skills. If the request asks how evidence changes or qualifies a thesis but does not ask for an investment choice or explicit prior-Decision review, keep it Research. When the user explicitly excludes an allocation or investment judgment, never create a Decision.