Our product releases are flying out the door, but my personal time is now the limiting factor: I cannot be the bottleneck for documentation. AI makes mistakes and misses context right now, but the knowledge it has can automate the PRD skeleton from a high level vision, problem or goal. This frees up a lot of time we spend staring at that first blank page, opening the floodgates for our high-leverage strategic judgement to spew out.
The speed of execution is currently our greatest competitive advantage. For the last five months, my team has successfully shipped features to our enterprise customers at an incredible velocity. Our features are **sales-led**, our releases are consistent, and the positive momentum is clear in our company’s growth metrics. The model is clearly working.
My responsibility is to look ahead and identify constraints before they seize up the whole machine. Right now, the bottleneck is my time. Specifically, the hours consumed writing detailed, yet repetitive, Product Requirements Documents. A PRD is the foundational truth for a feature; it must be rigorous, especially in a high-stakes, live broadcast environment. Yet, crafting the boilerplate, the background, and the non-technical context is a low-leverage activity that severely limits my strategic capacity.
The constraint of the creative process
I have always found that constraint forces a higher calibre of output. For instance, I once undertook a years-long, **self-imposed musical constraint**—a previous personal challenge I documented on this blog. What seemed restrictive was, in fact, liberating; it eliminated decision fatigue and sharpened my focus, leading to one of my most productive periods by concentrating my mental energy. We must view the current state of **generative AI** through the same lens of productive constraint.
It is March 2023: generative AI is public, powerful, and rapidly improving. Yet, we must be brutally honest about its limitations today. It is prone to **hallucinations**, it invents facts, and its training data has recency limits. In short, it is not trustworthy enough to be the sole author of an enterprise PRD.
This is the key strategic insight: its flaws define its role. We cannot trust it with the strategic **truth**, but we can trust it with the structural **process**. Its purpose is to be the perfect automaton, handling the repetitive, process-driven tasks that consume our valuable time.
Automating the scaffolding, not the strategy
My team and I have begun using generative AI exclusively for the **scaffolding** of a PRD. I am not asking the tool to define the strategy; I am asking it to define the structure to contain my strategy. I feed the AI a short, bullet-point outline that details the high-level vision, the core problem, and the desired approach. This input is already verified: it is the output of my strategic customer, P&L, and sales analysis.
The AI’s task is to instantly transform those critical few points into a fully formatted document skeleton: correct headings, professional tone, and all the necessary introductory and transitional text. This is a massive time-saver. It removes the hours we spend staring at the first blank page, opening the floodgates for our **high-leverage strategic judgement** to spew out. My function immediately shifts from content *creator* to strategic *editor* and *verifier*.
Focusing on the 20 per cent that matters
This deliberate automation of the 80 per cent allows me to spend my scarce time on the crucial 20 per cent of the document; the content that requires human strategic judgement and domain expertise. This is the content that actually ships customer value:
* **P&L and Success Metrics:** The AI cannot calculate the projected profit and loss impact of a feature or determine the correct **Key Performance Indicators** for our specific business goals; that is solely my strategic judgement.
* **Core Customer Stories:** Meticulously editing the generated user stories to ensure they authentically capture the precise voice and critical needs of our specific enterprise customer base, not a generic persona.
* **Architectural Judgement:** Adding the high-level technical dependencies and constraints that only a Product Leader familiar with the existing code base and live system architecture can provide.
The only way a Product Leader scales their impact is by ruthlessly protecting their time for judgement. By giving the AI a clear, constrained role in process automation, we have eliminated a significant bottleneck and increased our overall throughput. We are not just writing documents faster; we are deploying strategic thought faster.
Dear AI, you’re not saving me, you’re making it worse! The docs you produce are too generic and miss all the tribal knowledge. I spend more time editing your rubbish output to sound specific than I would have just writing it myself. Unless you can pull context from Slack, Jira, and the codebase, stick to generating boilerplate!
It’s so frustrating, innit? We’ve set up AI to do the easy stuff, but the stuff you actually need written down—the subtle trade-offs, the weird dependencies—it misses every time. I’m still manually writing the architecture decision records (ADRs) because no bot can capture the human element of why we chose the solution.