Episode 10: We built a Twitter AI agent that writes, learns, and schedules your posts

We’re back after a short summer break! Can taking a full month of June be called a short break? Yes, if you do it right.

Shoutout to one of our mutual best friends for throwing the wedding of a lifetime. The good old days are happening now, besties. The whole point is to be there for it.

But now we’re well rested and itching to show you what we’ve been building with Claude recently.

In the spirit of summering (summering (verb): the deliberate act of having the best possible time during the summer), we’re both trying to be as efficient as possible with many many items on our to-do list.

Barbara’s to-do list has “Start posting on X to see if it could be a valid lead acquisition channel for you”. Does she have additional time in the week to devote to this experiment? No. But we’ll make it happen anyway, and we’ll do it well.

I built a full Twitter agent in Claude Code that sources content from our existing Two girls, one Claude event transcripts, checks AI news, prompts me with targeted questions in Slack during the week, and then writes a batch of posts every Sunday and schedules them as drafts in Buffer. End-to-end.

The agent runs on three scheduled routines:

1. Monday, Wednesday, Friday at 9am: Claude sends me a targeted prompt in my Slack channel called #twitter-agent. Not “how’s your week going” or “what challenge did you overcome today.” Those are too vague and difficult to answer. The prompts need to be specific enough that I can sit down and fire off a quick response in 30 seconds. Something like: “Fable just came back. Were there workflows you had to abandon when it disappeared? What are you switching back to?”
   

2. Checks relevant news daily: Before sending each prompt, the agent checks AI news for anything relevant to what I actually use. I told it specifically: I use Claude, I care about design tools connected to AI (Figma), video editing, and marketing workflows. Don’t resurface every Mid Journey update or random model announcement. If it’s not something I’m actively working with, why would I be talking about it?
   

3. Sunday at 10am: The agent pulls my Slack responses from the week, reviews the latest transcripts from our show, checks for major AI news, and writes seven Twitter posts. Those posts go directly into Buffer as drafts so I can sit down with my coffee, review them, tweak whatever needs tweaking, add images if I want to, and schedule them for the week.

I’m currently drinking my morning coffee and editing the batch of posts that automatically landed in my Buffer account. I’ll edit or expand on certain parts, add more context, add relevant photos.

If you want to make the process actually end-to-end, you can just tell Claude to schedule the posts following your best-performing times (Buffer gives you that already), and the posts will go live.

I prefer to have the posts in my drafts so I can edit each post before scheduling. This also helps the agent learn what I like and don’t like, and improve over time.

I did not want this agent pulling from random sources. Everything it writes is grounded in content that already exists from me:

• Every transcript from our Two Girls, One Claude episodes. Bojana and I are labeled as different speakers in each transcript, so the agent knows what I said versus what she said and won’t accidentally attribute her opinions to me.

• Every podcast I’ve appeared on as a guest. E.g. my episode from Marketing Against the Grain by Hubspot.

• My existing writing style and tone of voice document (built for LinkedIn, but transferable).

• And five Twitter reference accounts that I follow and whose format I like. I gave the agent those accounts to learn what Twitter content looks like as a platform, not to copy their voice. The writing style comes from my context docs. The format comes from those reference accounts.

First tweet from the batch (follow me to keep me accountable):

Barbara ✸@brbbzzz

Me while traveling: I just really need to get back to my routine, I feel so behind on everything Me as soon as I'm back in the said routine: we should go to the Dolomites right now

5:28 PM · Jul 6, 2026 · 1 View

Bojana asked whether I’m going to categorize the types of AI news I want the agent to track.

I hadn’t thought about it at all, and she’s right. “AI” is massive. New tools, economics, ethics, new workflows for lead gen, for GTM, for outbound, for research.

Without categories, the agent will just resurface whatever is trending on any given day, and I’ll end up getting prompted about things I don’t care about.

I didn’t add categories during the build because I wanted to see how the first test run goes. But I already know this is going to need refining.

If the Monday prompt asks me what I think about the latest Gemini update, I’m going to ignore it, and that’s a wasted prompt. Always be more specific.

The whole system routes through Buffer’s API via Claude’s connectors. I set it up during the episode, connected my Twitter account, and told it to push the posts as drafts.

When it said “all seven posts are now in Buffer,” I opened Buffer and saw nothing. Classic.

I told Claude it was lying and it didn’t work. Turns out Buffer’s calendar view doesn’t show drafts for some reason, but the list view does. All seven were there, just like Claude said.

My bad.

Buffer’s free plan limits you to ten scheduled posts, which is enough for what I’m doing. If this becomes a real channel, I’ll upgrade, but for testing it’s fine.

The writing style was not great. Here’s an example post it pulled from one of our previous live event transcripts:

“I do everything with Claude now, so I think he knows everything there is about me. I made a soul document for it this week and I thought it was gonna be a few cute questions, instead it was like, I see that you have anger issues, so what should we do with those?”

Not the worst, but not good either. As expected, the writing style will take refining. But all good agents do. Don’t expect greatness on the first try (with humans or AI).

The other issue: every post was roughly the same length. The skill needs format variation built in, and I’ll be adding that after I review the first full weekly cycle.

Without the Slack prompts, this is just another AI content generator pulling from old transcripts. The prompts are what keep the content current and personal.

Say Anthropic drops a major update on a Tuesday. The agent catches it in its AI news scan, sends me a Slack message asking what I changed in my workflows because of it, and I respond in 30 seconds. On Sunday, that response becomes a post that reflects what I’m actually doing, not what the internet thinks about the announcement.

I’ve used a version of this Slack prompt workflow for clients before. It dramatically improves the quality of AI-generated content because the source material is fresh and personal instead of recycled from months-old context docs.

I need to add content categories for the AI news tracking (Bojana’s suggestion).

I need to build in format variation so every post isn’t the same length.

And I need to run it for a full week to see what the Monday Slack prompt looks like and whether my responses are detailed enough to produce good posts.

If the channel works, I transition to writing more of the content myself and using the agent as a drafting assistant. If it doesn’t, I saved myself months of trying to force a platform that wasn’t going to work for me.

Download my MD file here. I scrubbed them of personal info that make up my own AI agent - add your own info, context docs and account inspiration, and it will be good to go! You’ll be taken through a setup checklist after you import the md file.

Bojana’s team at HeyReach just hit $16M in ARR! And this community is the first to hear about it. We took 10 minutes at the beginning of our event to chat about startup life and how they managed to build and scale such an exceptional team. HeyReach has one of the best company cultures that has survived rapid scale and hiring.

Bojana joined about a year and a half ago, when the company was just about $3M ARR. Her advice on finding startups to join: stop trying to analyze runway and TAM. Find people you enjoy working with, whose values match yours, and who you’d want to stay with even if the company failed.

We’ve heard yall loud and clear. The live build sessions (with monthly guest appearances) are here to stay. But we needed some help to keep up with the content.

We’ve just onboarded a video editor to help us improve the quality of our videos so they’re easier to follow and more useful for everyone watching.

She’s mega talented, has an adorable pug named Amigo, and is officially our Gen Z correspondent that keeps our Millennial energy in check.

We’re just getting started, and we’re getting better. Pinky promise 🤞

Us because we have so many good AI workflows to show you all this summer ✨🔥

---

• Buffer (free plan, allows for up to ten scheduled posts/week): for scheduling Twitter drafts

• Whisper Flow: for voice-recording the initial agent requirements as a transcript

• Slack: for weekly prompts to enrich weekly content batches

• Claude: for literally everything

Honorable mentions:

• Summer Fridays Jet Lag Mask: for your face after five weddings in a month (obviously not sponsored, but should be)

• Snail mucin mask: for that extra moisture after flights

Next up is a guest episode! We’re diving into GEO/AIO, and we needed an expert in that space.

Nicolás Bignú from GEOMetrics is joining us to walk us through the latest in GEO and AEO, how to get your brand mentioned by AI, and show us his best AI workflows around that. If you work in content, SEO, or marketing and you’ve been hearing “generative engine optimization” everywhere without a clear explanation of what to do about it, this is for you.

RSVP for the July 16 episode