My post-conference playbook

I collected 4,346 LinkedIn posts about Dreamforce and 2,789 about HubSpot’s UNBOUND.

I posted the dashboard and my three takeaways on LinkedIn today.

But this kind of post-event social media data is useful beyond pretty charts. For any event or conference, you can use this playbook to build a target audience (ABM), find unique messaging angles (PMM), and surface market insights (strategy).

The tricky part here is the data collection. That’s where the playbook starts.

The short answer is to use Apify + your favorite coding agent. But you probably won’t get good coverage with a one-shot prompt. I’ve been playing with this and found a few things that really work.

I share the build process and lessons below. You can also download the full build kit here if you’d rather get right into the skills and scripts.

A bunch of B2B marketers made a big bet on Dreamforce or UNBOUND last week. Booths, dinners, lounges, flights, swag, and social media.

You might get badge scans, but that’s a fraction of the opportunity. The social posts are a bigger window into the event and the people following from home.

Think RevOps leaders discussing the pros and cons of Salesforce’s AI announcements, or demand gen marketers sharing what they learned at UNBOUND about AEO. Identifying the people engaging helps you segment your audience for highly targeted campaigns. Analyzing their posts helps you position your brand for that audience.

I used Codex to orchestrate the work, Apify to collect, and Jev to help classify the posts.

For the starter kit, you need an event, a few good seed posts, a spending limit, an Apify account, and Codex or Claude Code working in a local folder. The agent handles the scripts and saves the results.

Apify is the critical tool for this. Its “Actors” are ready-made programs for specific collection jobs. The four HarvestAPI Actors in the kit collect public posts without needing my LinkedIn session cookies or login. I connect my Apify account, and the agent can call them.

The actors for this play: post search, company posts, profile posts, and profile reactions.

One important note if you’re using the kit: use Codex or Claude Code, not standard chat or Cowork, and select “Work Locally.” IMO this is always the best way to work.

The tricky part is discovering all the social posts that truly relate to the event.

Searching the event name gave me a starting sample. But people use different hashtags, talk about a session without naming the event, or post from a company page that search misses.

So I started following the clues in the posts. Someone who posts repeatedly about an event is worth a closer look. Collect their other posts. Then look at the posts they’ve reacted to. That can uncover more authors, company pages, and language worth searching.

Two routes were useful in this build: pulling company pages directly, and following the posts and reactions of active people outside the host company.

Pulling every liker’s own posts was much weaker. Reacting to one event post doesn’t mean someone is writing about it. I got more from people already showing repeated interest.

This is the centerpiece of the kit. The agent collects a small batch, checks what it added, and uses those discoveries to choose its next move. Agree on the goal and spending limit up front so it can keep running without asking you about every batch.

I want to know how many new, relevant posts a round added, what it cost, and where it suggests looking next. A batch of 100 posts we already have adds zero.

When a route gets thin, change tactics. Two weak searches don’t tell you that the company pages or active posters are exhausted. Keep following useful leads within the budget, and save progress so the next session can pick up the work.

Relevance takes judgment, too. A post mentioning Salesforce isn’t necessarily about Dreamforce. Keep uncertain posts separate and read the source text behind any conclusion.

You end up with a collected sample, not all of LinkedIn. An optional reference set of known posts can help check what you recovered. Keep it aside during discovery so it remains a useful test.

I used the hot new model Jev from TypeSafe for the first time in this build. Everyone on X is saying it’s a big deal. I used it for classification, which seems to be the main use case. It delivered on the cheap and fast promise. Cost was $1.35 for 192,645 classification decisions. That’s 27 questions across 7,135 event-post records, separate from collection costs.

The questions still need care. “Positive about the conference” and “positive about its product announcements” are different judgments. Tightening that distinction improved our labels.

The labels help me find patterns. I still used Codex and deterministic code for a lot of the analysis, but Jev classifications were a good ingredient.

Let’s say I want to use my Dreamforce and UNBOUND analysis to grow this newsletter. Here’s what I’d do.

That’s the experiment I have in mind for Part 2.

Should I run this and report back?

I read all replies