Pragyan Tripathi

AI, databases, system architecture, Clojure. I write about the choices that hold up in production - and the ones that don't.

[Pragyan Tripathi @pragyansorbit]

@paulg I tried something similar for a product aimed at marketers and found a name I really liked. Couldn't get a good .com, though. Thanks for sharing the trick.

[Pragyan Tripathi @pragyansorbit]

@arvidkahl @podscanfm @audiohook Congrats, Arvid! I've been following you and listening to your podcast for quite some time. Really glad to see this happen for you.

[Pragyan Tripathi @pragyansorbit]

@jasonfried Read the postmortem. Curious how traffic is split across the four datacenters. Are customers tied to a region?

[Pragyan Tripathi @pragyansorbit]

@yongfook Curious what pushed you to do it now. Larger customers asking?

[Pragyan Tripathi @pragyansorbit]

@arvidkahl The number of 'and MCP' replies here 😅 Curious how you're handling permissions between the API and MCP layer in your own products.

[Pragyan Tripathi @pragyansorbit]

@thdxr What kind of workload actually needs the expensive token burn to compete? A lot of the replies here seem to be getting far on subscriptions.

[Pragyan Tripathi @pragyansorbit]

@dhh I can get pretty far from a prompt to a working app. It's the deploy, operate and grow part that makes me wonder how much of the stack can really disappear.

[Pragyan Tripathi @pragyansorbit]

@lika_build Exactly. Once a run can change state, retries need more than 'try the model again.' Give each step an idempotency key, record what actually committed, and let a resumed run skip completed effects. Otherwise a timeout can quietly become two writes.

[Pragyan Tripathi @pragyansorbit]

@Postgres_Daily Good default for teams already on Postgres. The win isn't just one fewer database: you can join vector matches to permissions and business data in the same query. I'd still benchmark recall and latency on the real workload before treating it as a universal replacement.

[Pragyan Tripathi @pragyansorbit]

@sutarabhiii The evaluation piece gets skipped too often. I'd test retrieval separately from generation: for a set of real questions, did the right source even make it into the top results? If not, a better prompt won't fix the answer.

[Pragyan Tripathi @pragyansorbit]

@pathikghugare I havent hit the quota for that. Once I asked how is it working it uses a proxy mechanism to send the same request claude code would send.

[Pragyan Tripathi @pragyansorbit]

@HarshitMaurya_ What worked for me are two things
I keep agent memory off and keep decisions in files that agent re-reads.
I also use compact command with a description everytime for the context I think must not be lost.
Somehow ChatGPT has been running multiple compacts without losing context

[Pragyan Tripathi @pragyansorbit]

@charliejhills Good list. I'd push back on 4.

Correcting it once doesn't mean it never repeats. What I kept hitting is that the correction lands as a new and the old one both stay, with nothing marking either as superseded. Retrieval surfaces whichever matches the query. Both are in there.

[Pragyan Tripathi @pragyansorbit]

@acatovicx https://t.co/ek6BU3tI0Y
most probably the regression could be the issue
Claude Code removed no time estimates rule in 2.1.111. It restored and strengthened the rule in 2.1.152
Either way I use CLAUDE.md to specify not to give estimates. It was a huge issue earlier for me as well.

[Pragyan Tripathi @pragyansorbit]

@acatovicx Claude Code is not measuring execution time.

It is the pattern it has learned from human documents. Those documents use sprints, days, and weeks.

Also there was a regression on system prompt of Claude Code as per reverse engineered prompt tracker.

[Pragyan Tripathi @pragyansorbit]

@HarryStebbings @pmarca so happy to see this your interviews have been great source of learning…

[Pragyan Tripathi @pragyansorbit]

@patrickbetdavid Be the best at whatever you do even if it is mowing the lawn.

[Pragyan Tripathi @pragyansorbit]

@brankopetric00 kamal on hetzner server cheapest and reliable

[Jesse Pujji @jspujji]

I recently spoke to a marketer who ran a $40M brand with just two designers and ONE AI process:

He gave me and my team a masterclass on using AI to scale marketing and creative.

He chains 7 different tools together for: ideation, image creation, video editing, and iterating based on performance.

I paid him 6-figures to build these systems for my companies.

Now, I’m giving them away for free.

Repost + comment “GA” to get the guide in your DMs.