2x Entrepreneur, 1x exit. Engineering is my chosen art form. Was nomad for a year and lived in diverse places. Currently, AI harnesses & DBs @Fin_AI (@Intercom)
Wrote this bit about how we collectively are creating more and more tech debt but not treating it as one! Skills, rules, commands and agents are great, until they aren’t.
https://t.co/jrVkbgwZ9s
For the past few months, I have been concerning myself with the problem of skills and agentic coding quality.
I wrote about what happens when hundreds of these things start piling up: 👇
Not an influencer but the way I keep my costs in check is pretty straightforward - /grill-me or /superpowers:brainstorming when I start working on something using Opus, come up with a doc (I like superpowers style of doc here), kill the session and implement that doc using sonnet in a fresh session. Review work on github directly.
While one is being implemented, move onto the next task in line. I am able to compete 3-4 low to medium complexity tasks in a day easily. Based on complexity, increase or decrease the number of parallel tasks.
I also have been trying to offload repetitive work to a CDE (that i prefer running on sonnet) so that I can scale my work as I go.
I think people forget the quality of life aspect a lot. I have a great example: when I was in Bangalore, I used to fall sick every 2 months or so (minor flu or something). I have been in Dublin for over a year and have managed to limit that to ~2 days last year. I don’t know exactly what it is but I bet the quality of food, air etc has a lot to do with it.
100%! It’s a sure shot way of spending 100s of $ without any meaningful impact. A lot of things anthropic recommends just generates them more money by increasing token burn. Not that none of those things are useful but you must understand why and when you should use them.
Another such feature is dynamic workflows. Sure it may have some use case for a highly sophisticated orchestration but 99% of the people do not need that (or any) level of orchestration. All of this feels very similar to the k8s hype - I fell for that hype train, this time I am being very cautious.
i hate how a few people will use a term and it will just become the "thing you should be doing."
99.9999% of you should in fact not being "looping" your agent
@TaoTeTrade Yeah. I remember thinking how bad it is going to get when that million dollar prize was launched for writing an article. They have created a monster that is actively bringing down the quality here.
I am getting a little tired of the AI discourse on Twitter. It feels like all the crypto grifters have now started generating long form slop and using X’s article writing growth hack to spread bullshit “Claude hacks” for reach. Wish there was a way to block tweets with articles in it.
For the past few months, I have been concerning myself with the problem of skills and agentic coding quality.
I wrote about what happens when hundreds of these things start piling up: 👇
If you read one thing today to understand how orgs should adopt AI, this should be it. Absolutely love the articulation of how we do “claude”. So many insights from our work on the Claude code adoption here!
@ponnappa I saw this yesterday and loved it! I don’t know how many times I have cancelled compaction before because I had no idea if it was just stuck or actually doing something!
Yeah sure but it’s an escape hatch in practice. “It was just my agent, I didn’t review it” becomes a plausible excuse and attribution kind of makes that excuse available. Without it, the text is just yours. I also know it’s a personal tick. I don’t like the cases where one can hide behind that excuse at all, that bothers me.
We do use Claude plugins backed by a GitHub repo. We use jamf to ship our marketplace and a base plugin by default. Everything else is inside a GitHub repo where there are role specific and team based plugins that people can install via the marketplace.
GitHub is still crucial for this entire process and everyone in intercom has a GitHub account now.