> Make sure your friends and users do not add booster comments in the thread. HN users are adept at picking up on those, they consider it spamming, and they will flame you for it. If a friend or a fan has something interesting to say, that's fine, but comments should not be promotional.
This has been such a gap in my process. Founder, 3+ years in on AI/agent first development. Lots of of hand rolled tools that worked but took away from core development. Tried Greenflash and it's giving me a read on my users that none of my existing analytics is, with actionable insights for both the product team and the sales/marketing teams. Highly recommend. Guys, keep up the great work!!!
Thanks Dave. Curious what your existing analytics didn't or couldn't show you that Greenflash does, and how did you go about hand-rolling tooling previously? Any lessons in the build-your-own portion that are worth sharing?
Good questions. So I use Amplitude which I love, but it's tracking mostly clicks. Great for portions of our product, but didn't give me a lot around the conversations.
Hand rolled was mostly looking at conversation logs and passing Claude those sometimes, but we never got to automating that, so it was "as we thought about it", not every day delivered to email and slack as you guys do. That's a big win. It's worth noting I actually still use Claude but WITH greenflash, so I point Claude at the Greenflash CLI and use it when I need to do even further research across different customers and correlate conversations with other signals in logs or product .. but the inbox of signals that's built in directly to Greenflash goes a really long way for us.
been a cool product to work with. we had raindrop before and it just took a bit too much work to get good signal out of it. greenflash has been a solid out of the box solution to get insights immediately. I've made product changes directly as a result of reading through a conversation that was highlighted to me. happy customer!
Greenflash reads and analyzes every conversation between your agents and users and surfaces where user experience actually breaks down, even when eng metrics look nominal. So far we've run over 12 million analyses across 330,000 conversations.
My co-founder and I live this every day. When we first started dogfooding Greenflash, it exposed us to issues we never thought about during development. One quick example: early on in our agent's lifecycle, our inbox flagged a number of conversations where users were struggling. None of our other observability tools alerted us, and they couldn't, because nothing was technically broken. No errors, no failed calls. A few of our users were asking, and re-asking our agent for the same answers, becoming increasingly frustrated over the course of their conversations. Every dashboard said those conversations succeeded. What Greenflash saw was that the users never got what they came for. Its analyses surfaced the root cause on their own: our agent was truncating its outputs due to technical limitations, so the core of what people asked for never came back. We assumed it was a subset of our users and segmented to find them. Turns out there wasn't a boundary, it was our entire user base. From there we used our agent skill to pull the analyses and pair them with Claude Code, and shipped a fix to our core logic. The fix was easy. But understanding something was going wrong would be borderline impossible without Greenflash.
Users will do crazy things with your product, and test cases can't really capture UX failures. With language as the interface, this happens at a far greater rate. This is the space we care about most: the gap between what your users want out of your product, and the experience your agent actually provides them.
Under the hood, Greenflash analyzes every message, tool call, and conversation your users have with your product, and those analyses surface the patterns worth looking at out of the box. The agent skills are the part we lean on hardest: you point them at our analyses of your conversations to pull evidence fast, or to dig across trends that aren't obvious on the surface, what's moving the needle, what a proposed fix would actually do. It's all reachable through the API too, so you can pull an analysis or hand it straight to Claude.
Today, we support automated prompt optimization based off user conversations. We surface concrete prompt changes built on our analyses, delivered via API and webhooks. The caveat is that it takes upfront wiring to set up properly, and you should still keep a person in the loop as a validation gate. That's how we run it internally, and it hums along quite well. That said, prompts are only one part of the picture, as the story above shows.
Pricing is simple: free up to 1M tokens/month, then $299/month plus $5 per additional million tokens.
We're focusing on product owners and folks responsible for driving revenue-generating products. We're betting the person who owns the product, not the person who owns the infrastructure, is the one who should own agent quality. We're not sure teams are structured for that yet, or whether agent UX is still falling into a gap nobody owns. If you've figured out who owns this on your team, tell us how. Either way: sign up, try it, and poke holes.
> Make sure your friends and users do not add booster comments in the thread. HN users are adept at picking up on those, they consider it spamming, and they will flame you for it. If a friend or a fan has something interesting to say, that's fine, but comments should not be promotional.
https://news.ycombinator.com/item?id=22336638 / Show HN Tips
Appreciate the feedback!
This has been such a gap in my process. Founder, 3+ years in on AI/agent first development. Lots of of hand rolled tools that worked but took away from core development. Tried Greenflash and it's giving me a read on my users that none of my existing analytics is, with actionable insights for both the product team and the sales/marketing teams. Highly recommend. Guys, keep up the great work!!!
Thanks Dave. Curious what your existing analytics didn't or couldn't show you that Greenflash does, and how did you go about hand-rolling tooling previously? Any lessons in the build-your-own portion that are worth sharing?
Good questions. So I use Amplitude which I love, but it's tracking mostly clicks. Great for portions of our product, but didn't give me a lot around the conversations.
Hand rolled was mostly looking at conversation logs and passing Claude those sometimes, but we never got to automating that, so it was "as we thought about it", not every day delivered to email and slack as you guys do. That's a big win. It's worth noting I actually still use Claude but WITH greenflash, so I point Claude at the Greenflash CLI and use it when I need to do even further research across different customers and correlate conversations with other signals in logs or product .. but the inbox of signals that's built in directly to Greenflash goes a really long way for us.
been a cool product to work with. we had raindrop before and it just took a bit too much work to get good signal out of it. greenflash has been a solid out of the box solution to get insights immediately. I've made product changes directly as a result of reading through a conversation that was highlighted to me. happy customer!
Curious about your experience with Raindrop - what about it took work to get value out of?
Greenflash reads and analyzes every conversation between your agents and users and surfaces where user experience actually breaks down, even when eng metrics look nominal. So far we've run over 12 million analyses across 330,000 conversations.
My co-founder and I live this every day. When we first started dogfooding Greenflash, it exposed us to issues we never thought about during development. One quick example: early on in our agent's lifecycle, our inbox flagged a number of conversations where users were struggling. None of our other observability tools alerted us, and they couldn't, because nothing was technically broken. No errors, no failed calls. A few of our users were asking, and re-asking our agent for the same answers, becoming increasingly frustrated over the course of their conversations. Every dashboard said those conversations succeeded. What Greenflash saw was that the users never got what they came for. Its analyses surfaced the root cause on their own: our agent was truncating its outputs due to technical limitations, so the core of what people asked for never came back. We assumed it was a subset of our users and segmented to find them. Turns out there wasn't a boundary, it was our entire user base. From there we used our agent skill to pull the analyses and pair them with Claude Code, and shipped a fix to our core logic. The fix was easy. But understanding something was going wrong would be borderline impossible without Greenflash.
Users will do crazy things with your product, and test cases can't really capture UX failures. With language as the interface, this happens at a far greater rate. This is the space we care about most: the gap between what your users want out of your product, and the experience your agent actually provides them.
Under the hood, Greenflash analyzes every message, tool call, and conversation your users have with your product, and those analyses surface the patterns worth looking at out of the box. The agent skills are the part we lean on hardest: you point them at our analyses of your conversations to pull evidence fast, or to dig across trends that aren't obvious on the surface, what's moving the needle, what a proposed fix would actually do. It's all reachable through the API too, so you can pull an analysis or hand it straight to Claude.
Today, we support automated prompt optimization based off user conversations. We surface concrete prompt changes built on our analyses, delivered via API and webhooks. The caveat is that it takes upfront wiring to set up properly, and you should still keep a person in the loop as a validation gate. That's how we run it internally, and it hums along quite well. That said, prompts are only one part of the picture, as the story above shows.
Pricing is simple: free up to 1M tokens/month, then $299/month plus $5 per additional million tokens.
We're focusing on product owners and folks responsible for driving revenue-generating products. We're betting the person who owns the product, not the person who owns the infrastructure, is the one who should own agent quality. We're not sure teams are structured for that yet, or whether agent UX is still falling into a gap nobody owns. If you've figured out who owns this on your team, tell us how. Either way: sign up, try it, and poke holes.