Wink Pings

20 Repos, a 10-Step Tutorial: Grok Bot Takes the Work Off Your Hands—and Leaves You the Permission Problem

Codez compiled a Grok Bot productivity list covering skills, team orchestration, memory, and tool integrations. The list is worth bookmarking, but even more worth copying is his method for drawing approval lines for the bot.

Codez posted a list on X of 20 GitHub repositories, themed around turning Grok Bot into a small team that can actually get work done. He sorted the repositories by Skill, Plan, Team, Memory, and Tools. The list is easy to bookmark; the hard part is the set of management actions that comes after it.

Grok Bot's setup is different from a chat window. In Codez's tutorial, each bot has its own cloud computer, can log in to tools you already use, click through interfaces like a human, and keep running after you close your laptop. In other words, you go from writing prompts to assigning work.

First, the list. Below is Codez's order, with all links preserved:

01 noodle: orchestrates agents with skills, written in Go.
02 mattpocock/skills: skills for real engineers, from his .agents directory, small and composable.
03 agent-skills: production-grade engineering skills covering definition, planning, building, verification, review, and release—26 in total.
04 superpowers: an agentic skills framework and development methodology.
05 Vercel agent-skills: Vercel's official collection of agent skills.
06 awesome-agent-skills: 1000+ curated skills, from both official teams and the community.
07 Gas Town: a multi-agent workspace manager.
08 gstack: 23 opinionated tools: CEO, designer, engineering manager, release manager.
09 BMAD Method: an agile AI development method for agent teams.
10 grokbot-field-notes: rules, playbooks, and bot roles left from 72 hours of hands-on Grok Bot setup.
11 Spec Kit: GitHub's spec-driven development toolkit.
12 Task Master: AI task management that can be plugged into agents.
13 AGENTS.md: a simple open format for coding agents.
14 Beads: a memory upgrade for coding agents.
15 mem0: a memory layer for AI agents.
16 Context7: real-time documentation for agents.
17 npx skills: an open agent skills tool.
18 Composio: 1000+ toolkits with auth and tool search.
19 GitHub MCP Server: GitHub's official MCP server.
20 Firecrawl: search and web scraping for agents.

Video here:

The list looks like a toolbox, but Codez's 10-step tutorial is more like a management manual. Several points hold true for both developers and ordinary users.

First, give the bot a job title, not just a one-line prompt. Names like Inbox Manager, Expense Manager, and Sales Outbound should come with a charter: what it owns, what counts as good, and under what circumstances it must stop and ask you. One line from the original is key: a bot with a cloud computer and login permissions can act while you sleep. Draw the boundaries in advance.

Second, connect tools once, but understand the blast radius. Connections are account-level: connect Gmail or GitHub once, and every bot you create later can use it. The upside is that the fifth bot can start working in seconds; the risk is that one connection's scope covers every bot you create in the future. During beta, connect fewer accounts than you think you need.

Third, use a handoff for login; don't paste passwords. The bot reaches a login wall in its own cloud browser, hands the screen over to you, and after you authenticate and click done, it continues from the same browser instance. The bot gets a session, not the password from your chat box.

Fourth, demonstrating once is easier than explaining ten times. Tasks you do every week, that span two or more tools, and that have stable steps are the best candidates for recording once and letting the bot learn.

Fifth, describe schedules and triggers in one sentence. A 7 a.m. briefing, a Friday pipeline summary, drafting a reply when a new email matches certain keywords. Codez says setting up a trigger-based routine in practice takes about two minutes.

Sixth, specialist bots are easier to manage than one do-everything bot. Separate bots have separate memory, context, and accountability. An Expense Manager that only watches receipts will be more reliable than a bot juggling reimbursements, recruiting, and outreach at the same time.

Seventh, put bots in a group chat and give them goals, not task lists. A task list means you've already done the decomposition and they're only executing. Goals let them divide the work themselves: one bot fixes a bug and opens a ticket, then hands the problem to another bot to debug.

Eighth, draw approval lines by reversibility. Drafting, archiving, tagging, summarizing, researching, preparing, and reconciling are reversible; let the bot do them and log what it did. Sending externally, spending money, making something public, deleting, and accepting terms are irreversible; stop and wait for you. Codez's example: Sales Outbound queued 36 drafts and sent 0.

Ninth, spend 15 minutes a week reviewing. Ask each routine: Did it run? Is the output right? Would I miss it if I deleted it? Have the bot report on itself, then manually spot-check one output. Automation rots quietly: site redesigns, process drift, and output turning into garbage can all happen while you sleep.

One commenter quipped in the replies that step one of these 20 repos is having the bot read 20 READMEs instead of doing work. It's a harsh line, but a useful reminder: a list is not something where more installed is always better.

image

Another commenter, Tyson Isham, suggested a more practical security move: create a read-only research repo, have an agent crawl each repository looking only for malicious instructions, mark their locations, then have a second agent review them. Skill files are also a supply chain; audit them before installing.

Codez's conclusion is that AI products used to put the model in a window while the work stayed in your hands. Grok Bot hands the computer, logins, memory, and time to the bot, and leaves the decisions to you. The change doesn't sound very technical, but it's a big shift in habits. You no longer ask "How should I phrase this sentence?" You ask "What exactly am I delegating, and where do its permissions end?"

The list is worth bookmarking. Pick one task that repeats every week, spans tools, and has stable steps; give the bot a job title; draw an approval line; then let it run for a week. Being able to delete one routine is more useful than installing all 20 repositories.

发布时间: 2026-11-19 14:30