Grok Bots Directory
Submit a Grok Bot

Run deep research with a Grok Bot: choose a pattern, demand sources, verify claims, schedule updates

A research bot is fast at gathering and weak at knowing when it is wrong. This guide shows how to choose a research pattern, set source and citation rules, verify the result before you rely on it, and turn a one-off question into a scheduled routine, checked against x.ai documentation on 2026-10-31.

Checked on

The playbook

What you'll get

A repeatable way to run a research job with a Grok Bot: a clear question, a pattern that fits it, a cited answer with a claim list, a verification pass you control, and optionally a routine that reports only what changed. If you want to compare research templates, read Best Grok Bots for Research first; this guide is about running the work.

The research patterns

Pick one before you start. Each pattern matches a template style in the directory; the names are examples, not the only choices.

Pattern Use it when Example template
Quick fact-check One claim, an answer today Researchy
Multi-source deep dive An open question that needs studies, documents and news Deep Dive
Evidence brief with a claim ledger Someone else will rely on the result Ethan
Primary records Court, spending, lobbying or archive material Deep Research by Jonah
Structured output and monitoring Lists, spreadsheets, competitor changes Deep Researcher
Budgeted multi-angle report One big question with a spending cap Swarm
Social pulse What people are saying lately last30days
Cross-model check A second and third opinion Multi-model consensus

When unsure, start with the quick fact-check and move up only if the answer is thin.

Prerequisites

  • A Grok Bot plan and the desktop app (step 2).
  • A written question with a decision attached: what will you do with the answer, and by when?
  • For plugin-based search routes, an account with that provider. Check its pricing yourself; we did not verify costs.
  • Time to open sources yourself in step 7.

Troubleshooting

The answer cites sources that do not say that

This is the main failure of research bots. Open each link. Delete any claim whose source does not support it, and ask the bot to redo that part from primary documents. Do not accept a citation you have not opened.

The bot writes with confidence and no dates

Add to the brief: every claim needs a source, publication date and event date; the confidence must be stated; and unknown is an acceptable answer. Several templates, such as Ethan, are built around this rule.

Two sources disagree

Ask the bot to list both with dates and say which is primary. If it cannot resolve the conflict, report it as unresolved in your own write-up.

A routine never ran or runs too often

A new routine waits for its next scheduled time. Check Settings, Bot, Timezone, then use Test. Schedules must be at least five minutes apart and a routine runs at most 200 times an hour. For research, daily or weekly is usually enough.

Paid search plugin bills creep up

Set a spending limit with the provider where it allows one, ask the bot to report its spend on each run, and switch big jobs to manual. Swarm shows a spend line by design.

An app shows Needs auth or Disconnected

Choose Authenticate (Authorize on phone) in Connect apps.

Limitations

We did not run any template, connect any plugin or run a routine for this draft. Steps marked untested are documented or read from a template but not run by us. The official docs we read list no general web-search connector; search plugins named below are ones that member templates list, and how to add them is not covered by those docs.

Sources

Checked on 2026-10-31:

  • Grok Bot: get started: plans, install, sign-in, creating a Bot.
  • Computer and apps: Connect apps, public X search, cloud browser, MCP servers by URL, no general web-search connector listed.
  • Skills, routines and automations: routine limits, Test, time zone, event triggers, approval advice.
  • Security: approval cards and what Bots can access.
  • Template pages in this directory for the patterns above.

Run a research job in 10 steps

  1. Write the question and the decision

    Write the question in one sentence, the decision it feeds, the deadline, and what counts as done (for example, "three sourced options with risks"). A vague question gets a long, unfalsifiable answer. Not run by us; this is a planning step.

  2. Confirm your plan and install the app

    As of 2026-10-31, per Grok Bot: get started, Grok Bot is available with Premium, Super or Ultra on The X Subscription, with paid Cursor plans, with a linked SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription, or with a seat on a self-serve Grok Business plan. Install the desktop app from x.ai/bot and choose Sign in. Accounts on Privacy Mode (Legacy) must switch to Privacy Mode first. Check the plans page before paying. Documented, not run by us.

  3. Choose a pattern and add a template

    Use the pattern table above. Open the template's page, read its skills and routines, and add it from the link there. The get-started doc does not describe importing a template, so the exact clicks are untested; if import is unclear, create an empty Bot (+, then Create a Bot) and paste the research rules from step 5. Start with one template.

  4. Decide your sources and connect only what you need

    As of 2026-10-31, per Computer and apps, Bots can search and read public X posts and trends without signing in, and can use the cloud browser for sites without an app. The X plugin (Connect apps, then sign in with X) adds your account data and needs X API credits. Notion and Google Drive can store notes. For search and extraction plugins, the member templates list Tavily, Exa, Firecrawl and Parallel (for example Swarm and Deep Researcher), and the docs say Bots can add MCP servers by URL; connecting them is untested here, and any paid service needs a budget. Prefer primary sources: filings, laws, datasets, original papers.

  5. Write the research brief with source and citation rules

    Paste the question from step 1, then these rules: every factual claim gets a link, a publication date and an event date; mark each claim as verified, estimated, inferred or unknown; quote exact text only from the source; say unknown instead of guessing; list contradictions with both sources; end with a short confidence statement and open questions. Do not allow social posts to count as proof. Ask for a claim list as a table so you can check it line by line. Untested for any specific template.

  6. Run it once and read the output as a draft

    Run the job in chat with a modest scope. Check it has a claim list, dates, and a clear separation between what is known and what is inferred. If it is a long run, tests and runs spend usage, so start small. If the output has none of these, tighten step 5 rather than asking for more detail.

  7. Verify the claims yourself

    Pick the five claims that most affect your decision and open the cited pages. Confirm the page exists, says what the bot claims and is dated as stated. Recompute any numbers from the raw data. For a quote, photo or video, search for the original. Then run a second pass in a fresh chat or with a verifier template: ask it to assign PASS, FIX or BLOCK to each claim with the evidence, as Ethan's claim verifier skill is designed to do. A second model run is not proof; your own check of the source is. Untested for any specific template.

  8. Cross-check what matters most

    For a high-stakes conclusion, ask the bot to find the strongest evidence against it, then repeat the question with a different pattern (for example a primary-records bot after a general one). Where templates offer it, a cross-model check can be added, but agreement between models is not evidence. Record what changed. Untested.

  9. Save the result and the open questions

    Save the final report, the claim list and a list of what is still unknown to a place you can find later, such as a Notion page or the Bot's files. Note the date, because research goes stale. Some templates keep a library or corrections log for this; do not rely on it unless you have read its instructions.

  10. Schedule recurring research with a routine

    As of 2026-10-31, per Skills, routines and automations, ask the owning Bot to create a routine and confirm: schedule and time zone, input source, expected result, approval boundary, and what happens when a source is missing. Ask it to report only what changed since the last run and to say "no change" otherwise. Set your time zone under Settings, Bot, Timezone. Schedules must be at least five minutes apart, a Bot can own up to 50 routines and each routine can run 200 times an hour; start with weekly. A new routine waits for its next scheduled time. Use Test to run it once; tests do real work, including calling connected tools, and spend usage. Require approval before the routine sends, posts or publishes anything. Documented, not run by us.

Inspired templates8

Templates mentioned in this guide. Each page lists its skills, routines and plugins. The Best Grok Bots for Research collection compares them side by side.

Researchy

Researchy

by Farzad

A research and fact-check desk that runs every pass on the latest Grok model with live web search. For anyone who needs sourced, dated claims instead of a trai…

  • Memories5
  • Skills1
  • Routines0
Deep Dive

Deep Dive

by Warren Wilson

A deep-research assistant that investigates any question across studies, primary documents, news, and social media, reads sources in full, explains contradicti…

  • Memories10
  • Skills22
  • Routines1
Ethan

Ethan

by JUMPERZ

Evidence-first specialist for deep research, claim verification, competitor intelligence, trend monitoring, and research quality audits. Turns a topic, company…

  • Memories0
  • Skills5
  • Routines0
Deep Research by Jonah

Deep Research by Jonah

by Jonah

A research bot that digs through primary sources like court records, laws, filings, spending, lobbying, papers, books and archives, checks them against each ot…

  • Memories0
  • Skills5
  • Routines0
last30days

last30days

by Matt

Research what people actually say about any topic in the last 30 days. Installs the latest last30days skill from GitHub, walks first-run setup (ScrapeCreators…

  • Memories8
  • Skills1
  • Routines0
Multi-model consensus

Multi-model consensus

by Austen

Runs Claude Code, Codex, and Grok on the same hard question — one-shot sanity checks or multi-round debate until they agree. Built for forecasts, critiques, PR…

  • Memories2
  • Skills1
  • Routines0

More playbooks

All playbooks