Trust me on this: paste this into Grok Bot before bed. It might change your life.

Overnight it finds 10 companies that need what you sell right now, the person to message at each one, and a first message ready for you to send.

Here's the prompt:

SpaceXAI's senior field engineer just showed his full Grok Bot setup. It works for any kind of work, and you can copy it today.

He runs a team of 4 bots:

> Roy, the orchestrator, hands out tasks and runs the loop

> Ed, the metrics bot, pulls data from X and the website

> Dex, the desig> Lou, the developer, is connected to GitHub and pushes the code

The observation: one bot doing everything is like giving a whole startup's work to one person. Each bot here has one job and its own computer, context and memory. In his words, "context is king."

The loop: a routine fires every 24 hours. Roy asks Ed what performs worst, has Dex replace it, and Lou ships the update. Nobody has to prompt it.

Smart tip: one prompt makes Grok Bot launch Cursor cloud agents on 3 models at once, and each one sends back screenshots and a screen recording. Every time a new model drops, he runs this fan-out to compare them on the same task.

The token tip: don't fan out right away. First get the prompt right on a model you already know, because a vague prompt gets read 3 different ways and burns tokens 3 times.

Full breakdown: elfin-kip-a11.notion.site/Grok-Bot-201-a…

Original video: youtube.com/watch?v=rOVomd…

Copying the setups of top engineers has never been this easy. SpaceXSI team shares theirs for free, and it's a real goldmine

Here's what I already use:

1. Nightly Audit Engineer researches your whole codebase every night at 4am and ships one cleanup PR per area.
Link: x.ai/bot/marketplac…

2. Tech Demos picks one library from your X bookmarks every weekday, asks for your OK, and builds a demo with a cloud agent. Every PR needs a screenshot and a video.
Link: x.ai/bot/marketplac…

3. tinkabot wraps any API into a plugin with MCP and skills. It starts with the data shape, builds the smallest version that works, proves it locally, and only then publishes.
Link: x.ai/bot/marketplac…

4. Import Bot moves your setup from Claude Cowork, Codex, ChatGPT or OpenClaw into Grok Bot and turns it into up to 5 expert bots. It never touches the old tools and scrubs secrets first.
Link: x.ai/bot/marketplac…

5. Company Docs Q&A answers product and how-to questions from your live docs, then Slack and Notion, and always cites the source.
Link: x.ai/bot/marketplac…

6. Critiquito turns a screenshot or Figma frame into ranked design fixes before you write the code. It never edits your files.
Link: x.ai/bot/marketplac…

7. Video Editor takes your raw talking-head footage and edits it into a short vertical video: best takes, captions of what you actually said, B-roll, motion graphics and music.
Link: x.ai/bot/marketplac…

8. Writing Bot is a writing partner for essays, emails and docs. It runs a structured revision workflow, so the draft gets clearer and better organized.
Link: x.ai/bot/marketplac…

7. pstack isn't a bot. It's the skill set behind two of these bots: about 60 skills for shipping code, open on GitHub in Cursor's official repo.
Link: github.com/cursor/plugins…

Every bot opens with its full instructions and memory, so you can see exactly how the team writes them.

This is the most detailed practical guide to Grok Bot I could write.

After more than a month of daily use, I can say it for sure: it's the best AI agent out there.

I've explored almost every feature, tested a lot of setups, and I have plenty to show. x.com/i/article/2113…

The creator of Claude Code says there's no secret to prompting. His whole method fits in 3 lines

Boris Cherny says prompt wording mattered a lot back in the Sonnet 3.5 days. Now Claude only needs to know 3 things:

1. What you want it to do.
2. How much effort to spend.
3. How ht thing.

He showed a real prompt that uses all 3. He asked Opus 5.5 for an interactive website about a podcast episode:

The goal: "narrate the entire episode, focus on the business wisdom."
The effort: "use lots of tokens."
The check: "iterate till you're proud of it."

He also named the look he wanted: "New Yorker or NYT data viz vibes."

No role-play or templates. One paragraph, the way you'd brief a coworker.

I made a prompt that rewrites any request using his method. It's on the image.

Shopify just published the internal setup behind its AI agent. The one that got 7x faster and cut token usage by 90 percent. Here is how it works.

The observation behind it: most of the agent's work is reading, not thinking. To write one store query, it reads about 40 tokens for amounts of context, little judgment, all billed at top rates.

So they split the work. The expensive model talks to the user and makes the plan. Reading-heavy jobs go to small open models trained for that one task. The expensive model only gets the result.

The clever part is how they keep the small models good. A separate AI judge scores every change, and they tuned it until it agreed with human experts: from 37.7 to 74.5, slightly above a frontier model's 73. Then every night, failed conversations get rerun and fixed, and the fixes become training data. 69% of the failures it tried got repaired.

One thing stays expensive: talking to the user and planning still runs on the frontier model.

You don't need to train models to copy this: give your coding agent a cheap subagent for reading files.

Source: shopify.engineering/open-model-tas…

My AI agent generated $1,680 with one LinkedIn message. Here's how

I gave Viktor one job: find brands on LinkedIn that need ads.

First, he rewrote my profile. Brands don't reply to an empty account, so the headline now says what problem I solve, and my best results are pinned : job listings. If a company is hiring a marketer, it has an ad budget but nobody to run it.

Viktor finds those companies, finds their head of marketing, and sends a short connection note. No pitch yet.

Once they accept, he studies their site and writes one message: a real problem in their ads and how to fix it.

That one message turned into $1,680, and the company has been my advertiser ever since.

I've prepared the full prompt I use. It's below, just fill in your details and paste it.

Prompt:

You are my LinkedIn sales manager. Your goal is to find companies that need what I offer, reach the person who makes the decision, and start a conversation that turns into a paid deal.

ABOUT ME (fill in):
- What I offer: [e.g. ad placements on my X account with 19k followers / design / paid ads management]
- My niche: [e.g. AI tools, eCommerce, mobile games]
- My best results with numbers: [e.g. "a post for brand X got 120k views and 900 clicks"]
- Target markets: [e.g. United States, Germany, UAE]
- Target industries: [e.g. Software Development, E-learning, Retail]
- My LinkedIn profile: [link]
- Price range: [e.g. $500 to $2,000 per placement]

STEP 1. FIX MY PROFILE
Review my profile and write:
- A new headline that names the problem I solve and for whom, not just my job title. Example: "Helping AI startups reach 19k builders on X".
- An About section, max 120 words: who I help, 2-3 results with real numbers from the info above, and one line inviting them to message me.
- What to pin in Featured: my 3 strongest cases, portfolio or reviews.
Do not invent results. Use only the numbers I gave you.

STEP 2. FIND COMPANIES WITH A BUDGET
Use the job listings signal: a company hiring for marketing has an ad budget but not enough people.
- Search LinkedIn Jobs for: "Digital Marketing Specialist", "Paid Ads Manager", "Growth Marketer", "Performance Marketing Manager", "Content Marketing Manager".
- Keep only companies in my target markets and industries.
- Skip agencies, recruiters, and companies with fewer than 10 employees.

STEP 3. FIND THE DECISION MAKER
For each company, open the People tab and find one person with one of these titles: CMO, Head of Marketing, Marketing Director, Head of Growth, User Acquisition Manager, Brand Manager.
Use search operators to narrow it down, for example:
("Head of Marketing" OR CMO OR "Marketing Director") NOT Assistant NOT Intern

STEP 4. GIVE ME A LIST
Show me 10 leads in a table: company, person, title, open job they posted, what they sell, and one sentence on why they fit my offer. Rank them by fit. Wait for me to pick.

STEP 5. CONNECTION NOTE
For each lead I pick, write a connection note under 250 characters. No pitch, no links. Mention something specific about their company and why I'd like to connect.
Example: "Hi [Name], saw [Company] is growing its marketing team. I work with [niche] brands on [what I do] and follow what you're building. Happy to connect!"

STEP 6. FIRST MESSAGE AFTER THEY ACCEPT
Study their website, their LinkedIn page, and their current ads or content. Find one real, specific problem: weak creatives, no retargeting, no presence where their audience is, outdated landing page, and so on.
Write one message under 80 words:
- Their name and the problem you found, stated in one line.
- How I'd fix it, in one line.
- One of my results as proof.
- A soft offer: a free short audit or a quick call. End with a simple question.
It must sound like a person, not a template. No "I hope this finds you well".

STEP 7. FOLLOW-UP
If there's no reply in 5 days, write one short follow-up that adds something useful (an idea or an example), not "just checking in". If there's still no reply, stop.

RULES
- Never send, connect, or post anything without my explicit approval.
- Max 20 connection requests a day so the account stays safe.
- Never invent facts about me or the company. If you're unsure, ask me.
- Every week, report: requests sent, accepted, replies, calls booked, deals closed, and which message worked best. Suggest one change for next week.

This free open-source tool cuts up to 65% of your AI coding agent's tokens. Here's how it works

It's called Caveman, and it has 114k stars on GitHub.

Your agent wastes most tokens reading long logs and files, then writing long answers.

Caveman shrinks that data before the makes the agent talk like a caveman: short sentences, same exact code.

The quality holds: in the author's test, the agent still got 18 of 18 answers right.

Just one command: run npx @caveman-ai/cli and keep working as usual. Your original files stay untouched.

Repo: github.com/juliusbrussee/…

Meta just shared the AI agent setup it uses internally to cut a 3–4 day task to under 1 hour. Here's how to copy it

Meta first gave one agent all the knowledge. It got confused and started making things up

So they split it in two layers

The boss agent. It does zero real work. hecks for duplicates, collects what's needed, runs safety checks and hands the job to the right expert. 6 of its 7 steps are fixed rules, no AI guessing

14 expert agents. Each knows only its own area and sees nothing else, so it can't mix up rules

How to copy it:

Keep your main agent as a router. If it's making real decisions, it's doing too much

Give every expert the same structure: what it knows, what it needs, its steps, common errors and fixes

Store knowledge in separate files that load only when needed. Update a file and every agent uses it next run

For a new task, reorder existing agents before building a new one. Meta shipped a whole new use case with zero new agent code

Result: under 1 hour instead of 3–4 days, and about 2/3 of changes land without major human edits

source: engineering.fb.com/2026/10/19/ml-…
The prompt based on the original blog is below. You can paste it into Claude Code in the project root directory. It will first analyze the tasks, then suggest a structure, and will only create the files after you approve them:

Your multi-agent setup might cost up to 5.5x the tokens and still work worse than one agent

UPenn ran a simple test. Same model, two setups. One agent alone vs a team of three: a planner that talks to the customer, an agent that works with the database, and a rules advisor. ts: airline, retail, telecom

The team lost in 12 of 13 tests, by 28 points on average. The agents spent their budget talking to each other and often ran out before finishing the job

But it doesn't mean one agent is always better. This was about agents that talk to users, where info comes turn by turn. For big autonomous tasks, other studies found teams can help, especially with weaker models

So the real question is which setup fits your task. I put a prompt in the comments that audits your setup and tells you what to keep, merge or cut

Paper: arxiv.org/pdf/2610.27655
Prompt:

Audit and fix the agent setup in this repo. Base decisions on "Does Multi-Agent Collaboration Pay When Agents Serve Users?" (arXiv 2610.27655) and the research it builds on:
- Agents that talk to users (info arrives turn by turn): a single agent beat a 3-agent team in 12 of 13 tests, by 28 points on average. Most of the loss came from the team burning its token budget on coordination and not finishing.
- Autonomous tasks that can be split up in advance (research, big batch or coding jobs): teams can help, mainly when a single agent is weak at the task. Earlier work put the turning point at about 45% single-agent success.
- If a planner can't see the tool schemas, most of its orders run wrong (only 25% ran as written; with schemas, 72%).
- Plan-and-review inside one agent's context worked as well as a single agent. Splitting the same steps across agents lost.
- Extra agents did not catch more wrong writes to data.

STEP 1. Understand what I need
Read the code, prompts, configs and CLAUDE.md. Then ask me up to 5 short questions about anything you can't infer: what the agent does, whether it talks to users, how often it fails today and how, budget or latency limits, and which actions are risky.

STEP 2. Map the current setup
List every agent and subagent: its role, the tools it can call, the tool schemas it can see, whether it talks to the user, its inputs and outputs. Draw a mermaid diagram.

STEP 3. Decide per flow: one agent or a team
For each flow, classify it as user-facing (turn by turn) or autonomous and decomposable. Estimate how well a single agent handles it today, from evals or logs if they exist. Recommend:
- one agent, for user-facing flows or tasks a single agent already does well
- a team only for autonomous tasks that split into independent parts, where a single agent is weak
For every extra agent you keep, name the concrete failure it prevents and the evidence for that.

STEP 4. Measure the overhead
For a typical task, count model calls and estimate tokens for the current setup vs your proposal. Mark handoffs, re-planning, duplicated prompts and summaries of summaries.

STEP 5. Fix what stays multi-agent
- Give every planner the full tool schemas (names, arguments, types), even if it can't call the tools.
- Keep the user-facing conversation in one agent.
- Pass structured outputs between agents, not prose summaries.

STEP 6. Fix what becomes one agent
Merge the extra roles into the main prompt as a two-step loop:
1) plan: restate the goal, quote the rules that apply, draft the next action
2) act: check the draft against the rules and the conversation, fix it, then run it

STEP 7. Guard risky actions
List every tool that writes, deletes, pays, sends or deploys. Add a direct check before each one: validate the arguments, confirm the rules allow it, and ask for confirmation if it can't be undone. Don't add an agent for this.

STEP 8. Budget and finish rule
Set a token or step budget per task. When about 15% is left, the agent stops exploring and wraps up. Log running out of budget separately from normal completion.

STEP 9. Report, then implement
Show me:
1. The current setup and the proposed setup (diagrams)
2. Problems ranked by impact, with file paths
3. The exact changes, including prompt text
4. An eval plan: 10–20 real tasks, old vs new at the same budget, comparing success rate, tokens and unfinished runs
Wait for my approval. Then make the changes, run the eval, and report the before and after numbers.

Never write a CV or a cover letter again

One guy built a setup that handles the whole job hunt for him. After it got him hired as an AI engineer, he gave it away for free

AI Job Search, 62k stars on GitHub

You give it your CV once. It finds jobs that fit you, writes a CV and ne, has a second AI critique them, and preps you for the interview.

His result: 69 applications, 20 interviews, 1 job

GitHub Repo: github.com/MadsLorentzen/…