avatar
Chrome

Following0Followers (X)22.67KPosts141Likes2.86K
avatar

this is f*cking wild... Amrita, a field engineer at SpaceXAI, just put out a free 1 hour courses on building a Grok Bot team that works as one, from 0% to 100%: > 2:52 - a chief of staff with three bots that also talk to each other > 19:53 - rules for exactly when a bot has to ask you first > 30:41 - a backend bot and a frontend bot agreeing on an API contract > 31:27 - two bots removed, the rest reroute the work themselves > 44:56 - bots that pull in only the teammate they need most paid courses on agents stop at a single bot, this one has a whole team coordinating before the hour is out what each bot should write down before the next one takes over, that's the guide below

avatar

Ria Goveas and Audrey Leung from SpaceXAI spent 55 min building a Grok Bot team from scratch, and showed the secret that tells each bot when to hand the work on worth jumping to: > 02:37 - onboarding a bot like a new hire instead of writing it a manual > 09:17 - Dr. Eggbot, the bot that interviews you and builds the next bot > 20:34 - three bots that take a meeting from notes to a drafted follow-up > 43:17 - when a task needs a new skill and when it needs a new bot > 45:49 - a bot that reviews every other bot each week for wasted tokens the first bot in the chain is allowed to stop, if the notes owe nobody anything, nothing moves to the next one if you want that same chain on your own bots, the file each one passes to the next is laid out in the article below

avatar

Nate Parrott, the Anthropic designer behind Claude Design, spent 41 min showing how he turns a phone note into a working prototype, and the hidden step that makes Claude work like a senior designer save these timestamps: > 00:00 - 5 tips that stop Claude's output from sliding into slop > 18:12 - the 4 tells that make a design look AI-made > 21:53 - a plain spreadsheet turned into a live dashboard > 25:34 - that dashboard rebuilt as a Claude Motion video > 30:28 - testing a design skill when a model can't judge taste once the taste is written into a skill, the dashboard and the motion video come out of the same short brief the article below hands that brief to Fable 5.2 and runs it as a full motion design studio

avatar

Jenna Nanpei, field growth manager at SpaceXAI, spent an hour showing the full five Grok Bot setup she runs it's the one her marketing work goes through, and each bot in it came out of a piece of her week she stopped doing by hand a team like this usually falls apart the moment a person has to re-explain the brief to the next bot here every step arrives already finished: > TrendSpotter scans the field and brings back what's actually moving > Win the Day turns that into the event plan > Luma Pages builds the registration page from that plan > Field Finder and the Outbound Agent carry it to the right list > the approval shows up in Slack before anything goes out she never explains the campaign twice, the brief is specific enough that the next bot just continues from it how to draw that same line for three bots and what to write down at every handover is in the article below

avatar

this is insane, OpenAI has introduced a new model that works inside your apps for you and runs 10x cheaper than their flagship GPT-6.1 Terra is the one built for computer use, it looks at what's on screen, moves through apps and writes like the heavier models do where it stands against their other models: > it beats GPT-5.6 Terra by a wide margin on Agents' Last Exam > it holds GPT-6 Astra's level while running 10x cheaper per token > the average task finishes over 3x quicker than on Astra or Sol > standard API pricing comes in at half of GPT-5.6 Terra > your ChatGPT Work and Codex allowance buys more runs now speed and price moved in the same direction here, so an agent can stay running all day instead of being rationed by the hour

avatar

Jediah Katz, agent harness engineer at SpaceXAI: "Grok Bot is our autonomous colleague who gets work done for you, not just coding, but all kinds of knowledge work" the thing calling itself a colleague is a coordinator bot that owns the work and farms it out to a fleet of workers so what you give it is a job with a boundary, not a prompt with steps, and the whole result depends on which job you pick the five jobs worth giving a colleague like this are in the guide below

avatar

in less than 16 min he rebuilt a Mac app and modded a game with REA, and neither program ever came with a line of source Claude Code does the reasoning here, Ghidra does the opening, and both targets are programs anyone can already have on disk how the two runs go: > the game gets decompiled first, then changed and played back > the clipboard app gets rebuilt from what the agent recovered > every answer has to come from what the disassembler actually returned > both builds get opened at the end, which is where the gaps appear a rebuild proves itself when someone opens it, never when it gets screenshotted the full setup, the engines it needs and the jobs worth starting with are in the article below

avatar

Morluto spent 57 min on how he actually takes apart software he never had the source code for and where that ends up he's the creator of REA, the repo that went from 20 stars to 300,000 in 12 days, so when he explains how it's used, it's worth reading what that run actually consists of: > source was never an option, so a disassembler goes on first > the agent drives it while the decompiling happens on screen > the recovered pieces get checked against the program while it's running > what gets written down at the end is the behavior, never the code itself the models everyone runs this on were never trained for any of it, which is why the tooling has to hand them the evidence where this method holds, where it stops, and the prompts for it are in the article below

avatar

THIS GUY RAN A REPO THAT LETS CLAUDE READ 6 KINDS OF FINISHED SOFTWARE, AND REBUILT A SCREEN THAT NEVER HAD SOURCE CODE Electron and JS apps, websites, .NET, APKs, firmware and native binaries all open up through REA, and it's MIT licensed his target was a finished build of his own iPhone app, without a single source file near it that same prompt without REA came back with a screen that looked right and behaved wrong inspect → evidence → spec → rebuild → compare, and the compare at the end is what settles it everything you need to run it, and the places this dead ends, is in the article below

avatar

this prompt turns Claude Code into a reverse engineer, a role people get paid $136,000 a year for REA + Opus 5.5 reads the app on your computer, the card holds the whole run in one prompt here's the guide: > step 1: install Claude Code and sign in on a paid plan, Codex, Cursor and OpenCode work too > step 2: run npx rea-agents setup, approve the plan it prints, it wires REA into the agent you picked > step 3: quit the agent completely and start it again, then ask it to list its REA tools > step 4: paste the prompt from the card, fill in the app and ONE feature you can say in a sentence > step 5: it stops after the findings and splits them into confirmed, inferred and unknown > step 6: then SPECmd gets written, the build comes from that file alone, and the tests go against the original the result: a feature you actually understand, and a build that proves itself instead of just looking right the install, the limits and five more prompts are in the article below