From grok-workflows
Multi-source web research that fans out searches, fetches sources, adversarially verifies each claim, and synthesizes a cited markdown report. Use when the user wants deep, fact-checked research on a topic, asks you to "research X thoroughly", wants a sourced report, or asks for /deep-research.
How this skill is triggered — by the user, by Claude, or both
Slash command
/grok-workflows:deep-researchThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Runs the bundled grok-workflows harness, which decomposes the question into
Runs the bundled grok-workflows harness, which decomposes the question into sub-queries, gathers sourced findings per sub-query in quarantined web agents, adversarially verifies every distinct claim (source quality, factual accuracy, recency), and has a trusted synthesis agent write a cited report. You do not re-implement any of this — you invoke the harness and act on its JSON.
/deep-research <question>
This skill bundles an entrypoint at <skill-dir>/scripts/run.mjs (thin delegator
to the centralized launcher logic in src/launcher.mjs; self-location still works
via the passed import.meta.url from the delegator) —
<skill-dir> is this skill's own directory, whose absolute path is announced in
your system context when the skill loads. Derive the entrypoint path from that
announced SKILL.md path and inline the absolute path into a single
run_terminal_cmd call (don't rely on the working directory or a shell variable).
The launcher locates its bundled harness itself (via the delegator), so no
repository path is needed:
node <skill-dir>/scripts/run.mjs "<question>"
The harness prints a single JSON object to stdout (progress logs go to stderr):
{
"question": "...",
"subqueries": ["...", "..."],
"report": "# ...markdown report with inline [Source](url) citations...",
"claims": [{ "claim": "...", "source": "...", "url": "...", "subquery": "..." }],
"dropped": 7,
"droppedBreakdown": { "duplicates": 2, "refuted": 4, "verifierFailed": 1 }
}
report to the user as the primary answer (it is already cited
markdown). If the user asked for a file, write report to a .md file.claims.length)
and how many were dropped (with the droppedBreakdown), so the user knows
nothing was silently capped.claims is empty, tell the user nothing survived verification and suggest
narrowing the question.Do not paraphrase away the inline citations — they are the point of this skill.
npx claudepluginhub lswank/grok-workflowsGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.