Thank you for outlining a problem that is not easily understood! We see it every day at Gluecharm. Specs and how they are done, and with who, are definitely the way forward. Probably the only one!
Mar, thank you for sharing your experience at Gluecharm! Danny and I have been seeing something similar. Yes, prototypes are often more expressive than PRDs, but they are often far away from what is needed to move to production. A great spec is the glue between what we're learning with prototypes and what's needed to generate production code.
I completely agree on this and we even both used the William Gibson quote. The AI acceleration is very uneven and that puts more emphasis on the parts that haven’t been dramatically accelerated (talking to customers, product discovery, making good strategic decisions, etc.)
It feels like we are already pretty close to that!
Neil, I'm seeing something similar int the vibe coding I'm doing. I'm often including little code snippets or data files in JSON to give the LLM as much detail as possible. For example, I'll have ChatGPT generate mock data in a JSON file and then use that as part of my prompt in v0.dev.
very close, of what we are also experiencing ourselves day to day. The only y difference I seen is that in practice there is something even more "core" than the specs, as the things is that depending on your background and understanding and role, you need different specs but at the same time is the same app, so there must be something under, sort of source of truth, we call it internally the Application Map.
Its been interesting to see a similar evolution in the AI coding tools. For example, Claude Code has an easy way (the claude.md file) to setup app-specific context that gets pulled into any work the code generation agent is doing.
As someone who hasn't coded for 20 years and then it was in FoxPro, I have really enjoyed the freedom of design that vibe coding brings about. My team sees my vibe coding (primarily PoC work) as a fly in their ointment. I have 2.5 observations.
1. "The prompt is the spec" works to an extent. The complete conversation with the platform is truly the spec. Could the entire evolutionary thread of the vibe be reduced to a single spec? Not likely because while some prompts are refinement, many are fixing bugs.
2. A day will come soon where there is no human readable code. There will be a layer of gobbltygoop AI code that sits just above machine code interpreter. Once that is done, comments become irrelevant, the spec is the pseudo code. What next? You get the AI to reverse engineer the spec for you after the code gels? And for what purpose?
2.5. Your apps - which are really the embodiment of your spec - become like a child or employee being trained. Does the training live in modified code or in the persona of the app? Does it even matter?
Thanks for writing this Ravi, it is a refreshing read, PM roles is definitely evolving, the more we double down on the things moving at "Human Speed" the best adjusted we will be to design the right spec(source code). It is rather the best time to focus on fundamentals of PMing. Customer empathy and Clarity of spec.
Thank you for outlining a problem that is not easily understood! We see it every day at Gluecharm. Specs and how they are done, and with who, are definitely the way forward. Probably the only one!
Mar, thank you for sharing your experience at Gluecharm! Danny and I have been seeing something similar. Yes, prototypes are often more expressive than PRDs, but they are often far away from what is needed to move to production. A great spec is the glue between what we're learning with prototypes and what's needed to generate production code.
Agree that specs are an interface to infrastructure.
But specs have never really been about code - they're a proxy for ideas, values, and priorities.
So yes, more people can go from idea to prototype faster. But like you said, the bookends - feedback, trade-offs, decisions - move at human speed.
IMO, that's not a tech problem, it's more of a systems and culture problem.
More here -> https://hchau.substack.com/p/ais-uneven-acceleration
I completely agree on this and we even both used the William Gibson quote. The AI acceleration is very uneven and that puts more emphasis on the parts that haven’t been dramatically accelerated (talking to customers, product discovery, making good strategic decisions, etc.)
It's funny because I was having this conversation with colleagues at work .
For the product I'm building on own via Replit the importance of specs is huge especially with building out agents.
I think the dream is that one day you can just focus on writing specs and then the agents take over from there :)
It feels like we are already pretty close to that!
Neil, I'm seeing something similar int the vibe coding I'm doing. I'm often including little code snippets or data files in JSON to give the LLM as much detail as possible. For example, I'll have ChatGPT generate mock data in a JSON file and then use that as part of my prompt in v0.dev.
very close, of what we are also experiencing ourselves day to day. The only y difference I seen is that in practice there is something even more "core" than the specs, as the things is that depending on your background and understanding and role, you need different specs but at the same time is the same app, so there must be something under, sort of source of truth, we call it internally the Application Map.
Its been interesting to see a similar evolution in the AI coding tools. For example, Claude Code has an easy way (the claude.md file) to setup app-specific context that gets pulled into any work the code generation agent is doing.
Would love to see the application map if you can share (and it’s not confidential, ofc)
https://blog.gluecharm.com/introducing-application-maps-the-missing-blueprint-c49f3c4ddf0b wrote this article to make it more clear
Loved seeing the example setup from Danny 👌🏻
Thanks Shubham :)
As someone who hasn't coded for 20 years and then it was in FoxPro, I have really enjoyed the freedom of design that vibe coding brings about. My team sees my vibe coding (primarily PoC work) as a fly in their ointment. I have 2.5 observations.
1. "The prompt is the spec" works to an extent. The complete conversation with the platform is truly the spec. Could the entire evolutionary thread of the vibe be reduced to a single spec? Not likely because while some prompts are refinement, many are fixing bugs.
2. A day will come soon where there is no human readable code. There will be a layer of gobbltygoop AI code that sits just above machine code interpreter. Once that is done, comments become irrelevant, the spec is the pseudo code. What next? You get the AI to reverse engineer the spec for you after the code gels? And for what purpose?
2.5. Your apps - which are really the embodiment of your spec - become like a child or employee being trained. Does the training live in modified code or in the persona of the app? Does it even matter?
Thanks for writing this Ravi, it is a refreshing read, PM roles is definitely evolving, the more we double down on the things moving at "Human Speed" the best adjusted we will be to design the right spec(source code). It is rather the best time to focus on fundamentals of PMing. Customer empathy and Clarity of spec.