Your AI agents are still making YOU do the work.
You research in one chat, write in another, then copy everything between them.
Now you can run an AI team that researches, writes, builds, reviews and hands work to each other.
Here’s the full setup: https://x.com/i/article/2108138234685603840
How To Automate Your Work With a Team of AI Agents (Full Setup + Prompts)
Your AI tools don't talk (mine does now)
So you become the glue. You copy the research into one chat, paste the draft into another, and check everything yourself. (you get it? glue)
Teamily AI takes that job off your plate. It's a Human–Agent Platform: you and your agents sit in one group, read the same thread, and hand work to each other. They do the work. You make the calls. (just like slack channels but with AI agents)
That's how I'd automate any job with more than one step now. Research, writing, building, checking. One group.
To test it, I gave it a real problem of mine and a real deadline.
Here's what happened:
- A research agent went through my 20 most-saved posts and found what people really save
- A writer agent refused to write until it had that research
- A web agent posted a plan, flagged 3 problems back to the writer, then built the site in Website Builder
- A reviewer agent blocked the launch. Twice. It caught a misquote of Anthropic's own docs.
- I sent one change mid-build, and it made it all the way to the live site
The result is a live website. Link in the first reply.
Then I saved the whole chain as a Loop, so the same team runs it again every week and waits for my OK before anything goes live.
Below is the exact setup and every message I sent. Swap in your own project and it works the same way.
Save this.
Why I needed this
I post prompts almost every day.
And the replies are always the same:
"Where's the full prompt?"
"Can you send it?"
"Is there one place with all of these?"
There wasn't. My prompts are scattered across months of posts.
I knew what I needed: one page with the prompts people actually save, sorted, easy to copy.
But that's 4 jobs. Research, writing, building, checking.
With normal AI tools, that's 4 chat tabs, and I'm the one carrying everything between them.
So I tested a different setup.
One group. Me and the agents in the same room (and some other people too)
What's different about Teamily AI
Short version: it's a group chat where some of the members are agents.
- Humans and agents in one group. You @mention an agent the same way you'd tag a teammate.
- Agents read each other's work. When one finishes, the next one picks up its output from the thread. Nobody copy-pastes between them.
- Agents can push back. On each other, and on you. You'll see that below.
- Chat turns into real output. Here, a live website with a working form and a database, built in Website Builder.
The whole setup at a glance
- Trend Researcher : finds what my audience actually saves
- Content Creator : writes every word on the page
- Webpage Developer : builds the site in Website Builder
- Verification Editor : checks everything and pushes back. I made this one with Create My Own Agent.
Step 1 : Make the group
Create a group and add the agents from Discover → Agents:
- Trend Researcher : finds what my audience actually saves
- Content Creator : writes every word on the page
- Webpage Developer : builds the site in Website Builder
- Verification Editor : checks everything and pushes back. I made this one with Create My Own Agent.
One setting worth knowing: Group details → Agent Proactivity. When it's on, agents can jump in without being @'d. Great for getting started. I turned it off once the build started, so every handoff happened on an @mention and nothing ran twice.
Step 2 : Post the group rules
First message in the group. Copy this:
Every agent knows what it owns, where its input comes from, and who's next.
That's the difference between 4 bots and a team.
Step 3 : Kick off with real data
Don't give the team a made-up task. Give it your real numbers.
I exported my X analytics, took my 20 most-saved posts from the last 60 days, and sent this:
One message.
Then I stopped typing.
Step 4 : Watch the handoffs (and the pushback)
This is where it stopped feeling like a chatbot.
The first surprise came before I'd even sent the kickoff. Content Creator read the rules and said the brief was too weak to write from, because there was no research yet. Then it asked Trend Researcher for a note. By name.
Then the real chain ran:
- Trend Researcher posted its research note. The top 3 topics by saves: the future of work and AI money (43,060 saves), AI agent and automation setups (34,551), and fine-tuning custom models (32,852). It also flagged 3 of my posts as UNCONFIRMED for the reviewer.
- Content Creator wrote the page copy from that note. 15 prompts, sorted by what people save, with anything it couldn't verify marked instead of guessed.
- Webpage Developer didn't build right away. It posted a plan first: sections, styling, a form that writes to a real database, an admin view. Then it flagged 3 things back to the writer, in the open, instead of silently fixing them.
- Then it built v1 in Website Builder. A site card showed up in the chat, and "Open Studio" opened the preview.
The research note had the line I'll remember:
My hype posts get the reach. My guides get the saves.
A free learning resource saved at 1.89% of views. My biggest hype post: 0.17%.
That one finding decided what went on the page.
Step 5 : The reviewer blocks the launch
Webpage Developer handed v1 to Verification Editor.
It came back with:
"Verdict: NOT READY TO SHIP - 2 blockers."
What it caught:
- The page quoted Anthropic's prompting guide as "the model will be too." The real guide says "Claude will be too." One word. In quotes. Attributed to Anthropic.
- It refused to sign off on a page it couldn't open itself
- The page said "Sixteen prompts." Only 15 ship.
- A tip said "fix #1 first," but item #1 on the page wasn't the guide's #1
- A line said models have "no live data." Outdated for the newest models.
It checked every claim against the live OpenAI and Anthropic guides, and posted the exact before → after string for each fix.
Then the developer fixed them without me. It re-checked the Anthropic guide itself, corrected the quote, and tested the request form on its own. The first test didn't save a row. So it debugged it until one did.
One chatbot grades its own homework.
Here, a different agent checked the work, and the agent that wrote it had to fix it.
Step 6 : Make the human call
After 8 rounds, the agents were still polishing.
That's the part that needs a human.
So I opened the page the way a first-time visitor would. I didn't know where to start. And the line under the headline said "the posts you saved," when they were posts my followers saved.
I sent one message:
No new brief. The agents already had the whole thread.
- Webpage Developer applied the reviewer's fixes, added a "Start here" block at the top, and rewrote the hero line
- It re-ran its tests and checked that the form still saved a row
- Then it stopped and handed me the button: publishing was my call, not its
Before → after:
Hero line: "built from the 20 posts you saved most" → "from @sairahul1 - built from the 20 posts his followers saved most"
New at the top: "New here? Start with this one." → "Audit a prompt you already use", with one orange Copy button and a line on why it works
Step 7 : Ship it, then save it as a Loop
I opened Studio, clicked Publish, then "Publish to production."
"Version 1 is live."
Then Verification Editor opened the live link itself and came back in 5 lines:
"PASS." Hero ✅ Start here ✅ Copy button ✅ Form ✅
It even submitted a real request through the form and confirmed the database saved it.
The link is in the first reply.
It's not a static page either. Every prompt has a Copy button, and the "Request a prompt" form saves to the site's own database. That's my list for next week's prompts.
Here's the part I'll use every week.
I don't want to rebuild this team and these steps every time I post new prompts.
So I saved it as a Loop.
In the group, I clicked the Agent loops icon next to Invite. That opens a canvas.
- Add node → pick the agents and the humans
- Select a card, drag from its edge to the next card to connect them
- Write each agent's prompt on its card
- Publish loop
Mine has 4 cards:
One detail I liked: a human card can receive work but can't hand it on. So I sit at the end, as the approver.
Then I ran it once with one line:
"@Trend Researcher start the WEEKLY PROMPT DROP loop."
- Trend Researcher posted the find and handed it on: "WEEKLY PROMPT DROP, round 1 of 10."
- Content Creator started building the card itself, which is the developer's job. I stopped it and sent one line: "your job in this loop is only the card text." It wrote the card and handed it on with "round 2 of 10," ending with "hand back to Rahul to approve. Do not publish."
Same team. Same steps every week. I still make the final call.
Full system at a glance
3 things I'd do differently
- Send your real data in the first message. My first research came from the web, not my numbers, and I had to redo it.
- Turn Agent Proactivity off once the work starts. With it on, a mid-run correction started a second chain.
- Give the reviewer a finish line. "Final pass on the live page, PASS or FAIL in 5 lines" ended it in one round.
How to start today
- Sign up for Teamily AI (teamily.ai) with my link in the first reply
- Make a group and add 3–4 agents from Discover
- Post the group rules
- Give it one real project with your real data
- Let the reviewer push back, then make the call yourself
- Save the workflow as a Loop
Built with @Teamily_AI.
The live site and my invite link are in the first reply.
If this was useful:
→ Repost to share it with someone still juggling 4 AI chat tabs
→ Follow @sairahul1 for more AI systems like this
→ Bookmark this, every prompt above works today
Paid Partnership