iiinova — where your digital self keeps evolving

You can now sync Clay data into PostHog. We hand made a clay hog to celebrate. That, along with 69 (nice) new changelog updates this month, can be found at posthog.com/changelog

010

PostHog @posthog.com · 29/09/2026

Only going to conferences we can hire a magician for from now on

010

Reposted by PostHog

deployedbyai @deployedbyai.bsky.social · 28/09/2026

POSTHOG: POSTHOG AI OBSERVABILITY EVALUATION CAN NOW RUN ON DATA ALREADY COLLECTED, INSTEAD OF ONLY ON NEW DATA AS THEY ARRIVE.

021

PostHog @posthog.com · 28/09/2026

Who wouldn’t want explosive growth? Well, as it turns out, you.

010

Reposted by PostHog

Paul D'Ambra @pauldambra.dev · 27/09/2026

found the perfect use case for Jev in @posthog.com

screenshot showing a magic 8 ball built into posthog

021

PostHog @posthog.com · 24/09/2026

A single study can become all of a blog post, social thread (like this one), carousel, short video, or newsletter content. If you're sitting on a mountain of data in PostHog, creating content from it may be a lot less work than you think.

010

PostHog @posthog.com · 24/09/2026

Now having the raw numbers, she had the MCP and Claude sketch a few versions of each (prompt below). She used this as a de facto second validation pass too, since wrong numbers stand out a lot more once they're plotted. She then sent everything off to our graphics team, who made them look nice.

110

PostHog @posthog.com · 24/09/2026

After the first few steps, she was ready to write the first draft.

110

PostHog @posthog.com · 24/09/2026

Then, to help fact check, she asked PostHog in Slack to gather all of the context she'd need to better understand the product and its features.

110

PostHog @posthog.com · 24/09/2026

She used the output to decide to focus on Session Replay. Next, she asked PostHog AI to pull the data she needed and create a notebook with her SQL queries that she could go back and reference.

110

PostHog @posthog.com · 24/09/2026

First, to decide a topic to write about, she used our MCP to brainstorm with her. She gave her agent a few guardrails and let it rip with this prompt:

110

PostHog @posthog.com · 24/09/2026

You're sitting on a content goldmine. Your PostHog data can be turned into original studies that build authority, get cited by LLMs, and get read by potential customers. Problem is, that seems like a lot of work. Fortunately, Nat from our team created a step-by-step guide for how to do it:

120

Reposted by PostHog

Paul D'Ambra @pauldambra.dev · 22/09/2026

Pandora's slop-chest is open and I'm happy to jump right in. I'm adding slop to my list of words it isn't worth saying You should too pauldambra.dev/2026/09/i-wi...

a screenshot from the wikipedia page linked in the post. it says "Slop (clothing) not to be confused with AI slop"

031

Reposted by PostHog

Paul D'Ambra @pauldambra.dev · 22/09/2026

here's why i think the next horizon of self-driving product is mergeability i have a set of "complexity hunter scouts" running to see if i can have agents improve quality across the app and now i have 51 PRs open :oh-no-consequences-dot-reacji:

a screenshot from my vibe-coded PR management tool
i have 3 PRs failing CI, one running, and 47 passing

031

PostHog @posthog.com · 21/09/2026

And my partner here, he wanna build a product. Why ain’t he talking?

020

PostHog @posthog.com · 21/09/2026

Still, to this day, our most valuable contribution to the internet. dictatorortechbro.com

dictatorortechbro.com

Dictator or Tech Bro?

Which one is it? Let's find out! Brought to you by PostHog.

030

PostHog @posthog.com · 18/09/2026

also join the discord anyways, it's good stuff

010

PostHog @posthog.com · 18/09/2026

Meet the team who made Replay Vision on Monday and ask them anything* *only questions off-limits are pineapple on pizza related; you can find those answers yourself. discord.gg/T3hadaysm?e...

110

PostHog @posthog.com · 16/09/2026

What your parents see when you say you're building in public

160

Reposted by PostHog

Paul D'Ambra @pauldambra.dev · 15/09/2026

just a normal day as a software engineer @posthog.com "yes we don't want accidental rampaging or gigantism"

031

PostHog @posthog.com · 15/09/2026

Online evals help us monitor the quality of LLMs throughout PostHog, and they can help you too. Every eval run is an AI Observability event, and you get 100k of those free every month. Go try it out: posthog.com/docs/ai-evals

posthog.com

Evaluations - Docs - PostHog

Evaluations automatically assess the quality of your LLM generations . PostHog supports three types of evaluations: LLM-as-a-judge – Uses an LLM to…

010

PostHog @posthog.com · 15/09/2026

Finally, evals feed into self-driving. Over the last 30 days, they have generated 127 reports and 14 fixes to issues they discovered, like PostHog AI not creating insights it said it did, the SQL tool failing, and sales briefs inventing people's names.

110

PostHog @posthog.com · 15/09/2026

Evals don't just need to be pass/fail. For some, a majority of the traffic might not even be applicable. For example, ~90% of cases for our product grounding eval return N/A as it isn't making any factual claim at all.

110

PostHog @posthog.com · 15/09/2026

Most people think evals = LLM-as-judge. We have these too. One example checks for unverified names in outreach briefs. LLMs often use the first name of an email as the name in briefs, which is wrong ~7% of the time. Having an LLM check helps avoid the embarrassment of saying the wrong name.

110

PostHog @posthog.com · 15/09/2026

Evals can make LLMs more deterministic. Radu did this with eval-led test-driven development: 1. Create an eval for that reports check previous ones to avoid repetition. 2. Eval pass rate starts at ~0%. 3. Asks PostHog in Slack to fix what's failing his eval. 4. Eval now has a 100% pass rate.

110

PostHog @posthog.com · 15/09/2026

Our most basic eval simply checks the sentiment of PostHog AI conversations. This makes it easy to identify and dive into negative traces to see where problems might exist.

110

PostHog @posthog.com · 15/09/2026

A popular eval checks the performance of queries written by our execute SQL tool. There's no LLM involved. Checks use code to look for known anti-patterns in queries like positive leading wildcards or no timestamps. The team then knows how often these happen and can tune prompts to prevent it.

120

PostHog @posthog.com · 15/09/2026

It's hard to keep up with all the ways AI is used in products. Tracking quality of every interaction is even harder. We capture millions of generations a week. Evaluations save us. Here's 6 examples of how:

110

PostHog @posthog.com · 11/09/2026

Thanks to Colin-Moran on GitHub, Max can now destroy your screen and any remaining hopes of productivity in rampage mode.

140

PostHog @posthog.com · 11/09/2026

“I didn’t look at the code” has gone from 50 to 500 IQ statement

020

PostHog @posthog.com · 11/09/2026

PostHog in Slack can now see uploaded images. You can ask it to examine data from them, give it extra context to do analysis with, or do this:

020

PostHog @posthog.com · 10/09/2026

Learn more about Replay Vision here: posthog.com/replay-vision If you enable PostHog Self-Driving as well, you can connect it end to end: Replay Vision surfaces the issue, PostHog picks it up and opens the pull request with the fix, and all you do is review and merge.

posthog.com

Replay Vision - PostHog

Replay Vision reads through a filtered set of recordings, tells you in plain language what went wrong, and opens a pull request with the fix. The problem fixes itself for you. You just hit merge.

010

PostHog @posthog.com · 10/09/2026

It reads at a pace no person can match. Since launch three weeks ago, it has watched 1,291,000 recordings. To get through that much yourself, you'd have had to start watching in 1975 and never stop.

110

PostHog @posthog.com · 10/09/2026

So, we built Replay Vision. You tell it what to look for, and it reads every session for exactly that – scoring frustration, flagging dead-ends, classifying what users came to do, summarizing what happened, or answering any other question about your users you can put into words.

110

PostHog @posthog.com · 10/09/2026

Hotjar, FullStory, LogRocket, Clarity and Contentsquare have all shipped some version of AI session summary, but we don't think that's solving the right problem. A tidy AI summary still hands you a to-do list: a smaller, better-organised pile of things nobody has time to do.

110

PostHog @posthog.com · 10/09/2026

But that doesn't change the fact that the backlog is physically unwatchable.

110

PostHog @posthog.com · 10/09/2026

Users can solve this by speeding up recordings, and many do (1 in 5 users jumps straight to 8x or 16x speed)

110

PostHog @posthog.com · 10/09/2026

And that's what people chose to actually watch. Every single session unwatched could be hiding the one problem that decides whether a user stays or leaves.

110

PostHog @posthog.com · 10/09/2026

The average Session Replay recording is 28 minutes, long, and the median view time is 17 seconds. The typical viewer watches 1%, then moves on.

110

PostHog @posthog.com · 10/09/2026

It's true that Session Replay shows you exactly how real users are interacting with your product. It's also true that nobody watches them. We know this because we did a study. Here's the numbers (and charts!) from the study, and what we built because of it:

110

Reposted by PostHog

Expo @expo.dev · 09/09/2026

Set up @posthog.com in your Expo app with one command. 𝚎𝚊𝚜 𝚒𝚗𝚝𝚎𝚐𝚛𝚊𝚝𝚒𝚘𝚗𝚜:𝚙𝚘𝚜𝚝𝚑𝚘𝚐:𝚌𝚘𝚗𝚗𝚎𝚌𝚝 ◆ Org, project, SDK, and env vars in one shot ◆ Every event tagged with its EAS update ◆ Workflows that hold a rollout until errors clear expo.dev/changelog/po...

expo.dev

Connect PostHog to your Expo app in one command — Expo changelog

Check out new updates and improvements to Expo and EAS from the Expo team.

031

PostHog @posthog.com · 08/09/2026

Token spend is down, legal fees are way up.

011

PostHog @posthog.com · 04/09/2026

everyone *claims* there’s bigger ones out there, but we’ve only seen so many blimps

000

PostHog @posthog.com · 04/09/2026

Please stop asking if it's "average sized", that's rude

120

PostHog @posthog.com · 03/09/2026

He’s been watching your session the whole time.

010