---

name: rule-research description: Use when a consumer wants to bootstrap stack-aware ESLint rules from LIVE documentation rather than ship pre-baked recipes. Triggers: rule research, research stack practices, generate eslint rule from docs, bootstrap rules for my stack, rules-research, rule-bootstrapping, no-head-element, исследовать практики стека, сгенерировать правило из документации.

<!-- @harness-posture: portable — prose research methodology over repo docs and the tiered trust rules; no harness primitives --> <!-- @dual-pair: rule-research-protocol --> <!-- spec: agents/rule-researcher.md -->

rule-research

Authoritative for: the Claude Code trigger for the rule-research protocol — a thin wrapper; the canonical, AI-agnostic protocol lives in agents/rule-researcher.md. NOT authoritative for: project goal — see README.md#why-this-exists; the protocol itself (detection, the L4-expressibility filter, provenance discipline, file contract) — see agents/rule-researcher.md.

This skill is a thin entry point. The full protocol — detect stack → research practices from canonical docs (context7 + deepwiki + a real fetch) → author a ResearchPlan + GenerateSelection filtered to L4-expressible rules → write two committed JSON files — is agents/rule-researcher.md. Follow it directly; this wrapper exists only so the protocol is reachable by a Claude Code trigger without duplicating its logic (single source of truth, per dual-implementation-discipline.md §7).

Provenance for every researched practice is gated by the tiered trust model — see .claude/rules/research-source-trust.md for the discipline (Tier 0 builtin / Tier 1 derived from a direct dependency's own metadata / Tier 2 consumer-acked) and agents/rule-researcher.md's "Trust tiers" table for the per-practice mechanics.

Run-moment provenance check {#research-run}

Before authoring a provenance entry against a live research source, re-check resolveAllowedSources/validateProvenance (packages/core/research/allowlist-resolver.ts) resolves it — a source that fails Tier 0/1/2 fails closed, it does not silently pass through.

To run: open agents/rule-researcher.md and execute its numbered protocol against the current project, then run ./setup --full to synthesize the researched rules.

Without this skill

An agent researching a stack's practices hand-authors ESLint rules ad-hoc — with unverified provenance and no L4-expressibility check — and ships inert "manual" rules that pass validation without a firing test, the discipline-theatre this project exists to eliminate. Fresh, stack-specific knowledge that lives only in live docs is missed because the agent works from stale training data.

With this skill

The agent follows agents/rule-researcher.md: it fetches canonical official docs (quoted provenance), proposes a rule only when the practice is a single-file forbid-selector with a single-token-diff pair, and writes two committed JSON files. The deterministic factory turns them into an executable rule + firing test — or records a research-only finding when a practice is not L4-expressible. Every shipped rule is provably non-vacuous.