Thank you guys for showing up on an unpredictable weather evening. So I'm going to talk, I named this talk Middle to Middle
because one of the Silicon Valley investors called Genitive AI Middle to Middle because it still needs humans at both the ends to solve things end to end. So I like that phrase and it sort of aligns with the topic today. So middle to middle.
Before we start, about me, I'm Bala. I'm a 3x founder. I spent a lot of time building and scaling tech at startups.
Currently building Voib. Voib is a private AI dictation and a no -bot meeting note taker for people who do confidential work. work.
So for today's talk, I'm going to pick up a feature that I built recently and use it as an example to talk through about my current workflow and how it changed in the last six months.
I think last time I gave a session that was on my workflow then, back then, and I tried to do a live demo last time i'm not making that mistake again now so yeah some context setting so i think these are
some things to take away that uh a software development life cycle has uh of these roles uh if you guys are uh from tech background it'll be familiar to you that uh the it starts with the product manager who's responsible for the spec the requirements on what to build right and then it goes to the architect who does the technical design on how system gets built right then using
that design we typically break down into tasks then give it to engineers who will build each of this module then QA engineers are testing these modules and DevOps engineers are working on automation to deploy and maintain it right so what changed uh is all the roles remain the same it's just that today it's just me and claude uh sharing these roles and getting things done
end to end right and uh what when what i mean by middle to middle is uh at the beginning i take the role of the product manager and architect and leave everything else to claude to do and i come back in the end to make sure things are built correct before shipping so i believe uh there
are two things here one is the role and second is uh thinking about it in compartments though i am doing uh these roles myself i think about these as independent steps okay and i work on uh one stage at a time so that it's clear and every stage has an output uh which becomes the input to the next stage so uh so so so we so we achieve what we want and i believe this framework can be
loosely adopted to non -tech uh tasks also for for example i use this for my copywriting uh copywriting website design emails all of this uh especially on things where i am not an expert on yeah wherever i can't be the expert i give that role also to claude hey you be the copywriter you are the designer you build it and then there is another reviewer at uh in the middle of the process who's reviewing that task yeah and this this also helps me learn uh the unknown things
uh with respect to what changed uh the big change is i stepped out of the middle what i mean by that is the the life cycle and the process remains the same however uh six months back uh once we build the design then we used to like break down the tasks then it'll be a lot of small conversations with cloud where hey go build this api in the back end now let's use this api to build this application right now let's test this right so this is how the the my collaboration
with cloud was six months back now uh cloud has become and every other ai has become so good at long horizon tasks executing things in loop reviewing its own work and fixing mistakes uh it's gotten and remembering the context all of that so i i fully trusted cloud to do things to do to build the entire thing in one go so let's take an example and go through this
so this is a feature we shipped recently previously VoIP had like a single plan so it was all hard -coded recently when we launched meeting note taker as a product we wanted to have a system to define and manage plans and also map features features in a config -driven way so we can add whatever pricing plan we want in an easily extendable way.
So, though it's not a user -facing feature, technically it was a little challenging. It touches multiple repos back in the app layer and the website.
And this is something I always ways do for a lot of the big projects I execute with cloud. These are in some way mapped to the roles that I showed before. These are sort of the artifacts that are the output at each stage.
As a product manager, the way it works is I am having a conversation with cloud, working through the spec on hey what is the objective of this feature and what are the boundary conditions what is the scope what's not in scope so this is like a long conversation and anyway Claude has the complete understanding of
the existing system so it it comes back with assumptions questions and all of that and after the end of the discussion this is the output we produce which is is like a spec document.
And I spent a lot of time here reviewing this work because in the entire lifecycle, this is the most important document, according to me.
So verifying all the critical terms, what we are defining with respect to this particular feature, these are the features, these are the plans we are going to map and how the customer journey is going to look and so this is what I verify and if you notice every line item is numbered this makes it easy
to have a back -and -forth discussion with Claude then once design is done so that's what I meant by compartment so this output is ready now to take it to to the next stage, which is the design stage.
Now I wear my technical architect hat and work on the technical design, where here the focus is about defining the non -functional aspects of how this feature should be built.
And you don't have to worry about the content. I'll just, I'm browsing it, skimming it through just to highlight some of the critical aspects.
So here it's about what is the design principles. Some of it is already defined in our project scope. So it knows what is our design principles and existing system architecture so that it doesn't derail from it.
But the typical challenge is making sure it doesn't over -engineer. Also, it doesn't simplify things too much. So that's my role in discussing the level of complexity we need for this feature.
And as part of this stage, I'm also defining the test plan and what are the things I expect to be tested as part of this project project to make sure we don't miss anything critical.
So the two outputs will be coming out of this stage. One is the design. Second is the test plan.
So I review the design and the test plan essentially at the user journey level to make sure we are taking care of all the scenarios so that's it my role is done now i hand it off to cloud where it's it's going to build
its own task list on uh how it is going to execute this project end to end so so as you could see b1 b2 b3 these are all the tasks cloud has created for this project so about like 30 40 tasks these would have been like at least 10 or 20 iterations before now it's but it's it's created a task it's also identified in which order to execute it builds it has
the test plan so it creates the test cases tests it make sure it the issues are fixed and moves on to the next task and completes things end -to -end and so So usually what I do is I leave my Mac in like a finite mode and let Claude do it while I'm away.
So that once it's done, it finally produces this document, which is I call it maintainer guide, because I'm going to maintain it. So, here the purpose of this document is to know how to deploy, how to maintain this application. This also becomes the resource for when we build the next feature so that we build on on top of this and in a way without disrupting any of our design and also our deployment practices. So this is how we keep every feature built in the same, following the same principles.
For critical features like this, where it's going to be deployed to production, I give the local environment full control to cloud, where it's building, testing, and deploying things locally. Once it finishes everything, it creates a maintainer guide and gives it to me, where where I go follow that instructions, deploy in my dev environment, do the sanity, and ship it to production.
So this is my new workflow. And I feel the idea of breaking the project into tasks, thinking in compartments, deciding which roles you can keep, which roles you want to give it to Claude, are things that we can potentially take away and use it for all the tasks.
And I think our responsibility still is required at both ends to make sure Claude is doing a good job.
Yeah, that's all. Thank you.
And like me, you can also talk to Claude. Now there is a Black Friday sale going on.
Questions?