unicodef1wn/grokbot-field-notes

Grok Bot Field Notes is a public collection of files, templates, and playbooks that three engineers from the xAI Grok Bot team extracted from a stunt build: they started with an empty repository and launched a product in 72 hours, live on stream, using their own agent platform. The repo is the distilled record of that run, packaged so anyone can copy the rules into their own coding agents.

Who made it. Three engineers from the xAI Grok Bot team: Roshan Sadanani (product), Lauren Tan (engineering), Matt Palmer (developer experience). The repo itself says it's "what I pulled out of those three days" — one person's distillation, though the README doesn't name which of them did the pulling.

What's inside. A set of documents that fall into a few buckets:

• Rules for your own agents. AGENTS.md is a single file of house rules; copy it into your repo root and any coding agent reading your repo will pick them up. ANTIPATTERNS.md lists 40 things that broke on air during the live build — each entry gives what broke, why, and the rule that came from it.
• Component libraries. agents/ holds the longer reference files that AGENTS.md points to (verification, orchestration, prompts). roster/ has 69 agent roles, one file per role, each describing what the role owns, what it doesn't, where it gets facts, what needs approval, and a paste-ready description. playbooks/ has nine role workshops (engineering, PM, founders, sales engineering, sales, SDR, support, post-sales, marketing) with the bot teams, workflows, prompts, and numbers. ports/ converts the roster into Cursor-specific rule files.
• Guides and references. guide/ contains two PDFs: the 24-page Grok Bot Guide by SpaceX Engineers (the story of the three days, mental model, failure log, economics) and the 14-page Grok Bot Marketing Guide (one campaign run across six bots, launch-day growth, distribution). reference/ has two short files — ECONOMICS.md (every cost and metric, each paired with the rule it supports) and PRODUCT.md (which Grok Bot features affect how you design a bot). notes/ has structured day-by-day notes backing every claim.
• Workflow helpers. .github/ contains a pull-request template that turns the verification loop into a form: reproduced, proof attached, scope, needs a human. The README explicitly lists "migrations, deploys, money and permissions" as actions that should keep a human gate no matter how well the agent loop is running.

Core philosophy. The README's one-paragraph summary captures the operating principles: Give each agent one narrow job and a name. Build a verification loop before building a second agent. Make an agent reproduce a bug before fixing it, and attach proof to everything. When an agent is wrong, write down the general principle, not the specific story. Audit routines weekly because "frequency is where the money goes." Keep a human gate on the dangerous actions.

License and usage. MIT — you can copy anything from this repo into your own.

Notable gaps (things the README does not claim). The repo does not describe how to use the xAI Grok Bot platform itself, does not include the product's source code, does not claim any performance benchmarks beyond the 72-hour build, and does not document current maintenance status or contributor guidelines beyond the license.