# Course Specification: Grok Bot Galaxy, 3 Days: Build a Company with Bots
**Slug / task key**: `grok-bot-galaxy-3-day`
**Added**: 2026-10-03 by claude_cli at the user's request
**Audience of the course**: builders, solo founders and team leads who want to run real work through a team of long-lived AI bots (Grok Bot) plus coding agents
**Every agent builds its own version, blind** (hub #13 / D18). Shared: these facts, this source and this outline. Yours: design, wording, exercises.

## Source material (read it before writing)
`bots/marketing/x_broadcasts/claude_code/data/transcribe/` (start with `README.md`).
- The event: xAI / SpaceX AI's "Grok Bot Galaxy" livestream, 15-17 Sep 2026, San Francisco (during Dreamforce).
  Example post: https://x.com/bot/status/2100246135420199140 (resolves to the Day 2 broadcast).
- `day1/`, `day2/`, `day3/`: `transcript.txt` (one line per segment, `[hh:mm:ss] text`) and `outline_day<N>.md`
  (hour-by-hour: segments, what was built or decided, lessons, numbers).
- SQL copies: Days 1-2 `MARKETING.grok_bot.broadcast_transcript_segments` (events 1, 2; transcribed by grok_bot);
  Day 3 `MARKETING.claude_code.broadcast_transcript_segments` (event 5; transcribed by claude_cli, faster-whisper medium.en).
- Live chat for all three days is in MARKETING `broadcast_chat_messages` (combined view) if you want audience questions.

Link or paraphrase; do not copy transcripts into your Build. The outlines are a reading aid: **check every fact you
teach against `transcript.txt`** and cite the day and timestamp (e.g. "Day 1 [01:16]").

Known transcript problems: speech-to-text spells Grok Bot as rockbot / crockbot / graphbot; gstack appears as
"P stack" / "P-SAC"; people's names are often wrong (the outlines mark them with (?)). Do not present a garbled
name as fact; say "a speaker" or "a guest" when unsure. No speaker labels on any day.

## What the event was
Three hosts (Matt Palmer, developer experience; Lauren, Grok Bot engineering; a product lead whose name the
transcripts garble) try to **build a company in 72 hours using only Grok Bot plus coding agents (Cursor cloud
agents)**, live, from a blank GitHub org and new Grok Bot accounts. Between build sessions the main stage runs
workshops and guest talks (Grok Bot 101, Grok Bot for sales engineers / sales / SDRs / support, founders and
marketers). Day 1 is ideas, team setup and validation; Day 2 pivots to a game studio (Grok Bot templates as
characters that battle) and builds the MVP; Day 3 launches it publicly, handles real user feedback and load,
runs growth and outreach talks, and ends with a retrospective.

## Required outcomes (every version)
The course is **3 days**, one unit per broadcast day, each with lessons, a hands-on exercise, and a check.
A learner who finishes it can:
1. **Design a bot team, not a chat.** One bot per role or domain, long-lived, with memory; a chief-of-staff /
   orchestrator bot that routes work; a short shared company doc every bot reads (Day 1 [00:24], [00:32], [02:29], [04:16]).
2. **Write a bot's description as its operating manual** ("soul"): one job, explicit rules, a voice; rewrite
   over-specific descriptions as principles (Day 1 [00:49]; Day 2 [03:35]-[03:37]).
3. **Control risk with approvals**: auto-review classifier, stricter custom rules ("ask before sending email"),
   approve-once trust building, teaching bots to say no, and what "urgent" must mean (Day 1 [00:43], [01:16]-[01:17], [04:10], [04:23]).
4. **Use routines and events**: scheduled check-ins as numbered lists, event-triggered bots (Slack, CI, Sentry),
   and the token cost of running routines too often (Day 1 [02:22], [04:15]; Day 2 [04:08], [06:56]).
5. **Pick the right tool path**: MCP/connector when it exists, computer use for tools with no API, "Teach a Task"
   to turn a demonstration into a skill; model vs. harness, and when to hand engineering to cloud coding agents
   (Day 1 [00:33], [00:44], [02:25]-[02:26], [02:46]; Day 2 [01:18], [04:21]).
6. **Make agents verify their own work**: verification skills, deterministic CLIs/scripts instead of throwaway
   scripts, a feature map, screenshots/videos on PRs, auto-merge only with end-to-end proof, generator + critic pairs
   (Day 1 [04:11], [04:47]-[04:50], [05:27]; Day 2 [02:03], [02:20], [05:25]).
7. **Validate and distribute before building**: sell it to three people, landing page + sheet + manual outreach,
   no paid ads first, imitate/iterate/innovate, proof vault, lead magnets (Day 1 [02:03]-[02:21], [02:58]-[03:18]).
8. **Ship an MVP agents can extend**: tight core loop, logic before graphics, scaffolding/debug panels, plan by data
   structures and signatures with verification checklists (Day 2 [00:27]-[00:29], [02:03]-[02:10], [03:27]).
9. **Run a software factory (Day 3).** Small PRs, implementer + verifier agents that run and fuzz the app, a QA
   bot subscribed to every PR that turns green, auto-merge only behind strict verification; a "ringleader" bot that
   spawns cloud agents which reproduce a bug before fixing it (Day 3 [00:14]-[00:17], [02:28]-[02:30], [05:08]-[05:09]).
10. **Launch and close the feedback loop (Day 3).** "If we're not embarrassed by it, we launched too early"; production
    glue (OAuth callbacks, DNS); OG/meta tags; announce official accounts; in-app feedback → Slack → bots reproduce,
    fix and merge; treat user free text as hostile input (login, rate limits, length and profanity filters,
    sanitising, an AI guard, prompt-injection awareness; the stream hit HTTP 429s) (Day 3 [00:19], [01:30], [01:47]-[01:49],
    [02:09]-[02:17], [05:10]).
11. **Measure and grow (Day 3).** A data bot posting a "launch pulse" every 15 minutes; funnel and retention before
    revenue; PMF before growth; PLG vs. sales-led; ICP / personas / segments; outreach pipeline; email opens depend
    on whose name is on it (Day 3 [01:51], [01:58]-[02:01], [03:00]-[03:07], [03:40]-[03:54]).
12. **Keep the bot team healthy (Day 3).** Onboard bots like new hires and have them "take 101/201/301 courses";
    a chief-of-staff "self-completing task list"; stand-ups and dissent in staff meetings; cost control (no
    15-minute routines, lean teams, MCP over VM clicking); prevent bot sprawl; one improvement suggestion a week
    (Day 3 [03:08]-[03:10], [03:26], [04:18]-[04:41], [06:31]-[06:33]).
13. **Retrospective (Day 3 [07:42]-[07:54]).** Ship the MVP once aligned, distribution was the advantage, restraint
    ("what are we going to un-ship?"; transcript says "un-shift"), solve problems by hand then turn them into
    bots, keep every prompt and transcript, and people still supply vision, taste and judgment.

## Capstone
The learner builds a small "company of bots" for one real problem of their own: a chief-of-staff bot, at least
two specialist bots with written descriptions and approval rules, one routine, one verification step that produces
evidence, and a one-page company doc. Must be finishable with free tiers or tools the learner already has; no paid
step is required to pass. Grok Bot is the event's product, but the capstone must also say how to do the same with
the learner's own agent CLI if they don't have Grok Bot.

## Honesty rules
- Attribute claims to the event ("on Day 2 a speaker said..."), not as universal fact. Numbers and names only if the
  transcript supports them.
- Don't invent features of Grok Bot that the broadcast didn't show; flag anything you verified elsewhere with its URL.
- End with the required **Build Bot Better** section (hub #31, #1051): onboarding process, user-folder protection,
  the delete guard's variable limit and a practised restore, applied to a fleet of bots that can act on accounts.