Product Discovery is the magic that allows small teams to deliver world-class products. As product engineering becomes more homogenized (Claude/Codex doing most of the work), product decisions make companies stand out.
As a PM, product discovery was where I found the most joy (and still do). Taking a complex problem that feels too big to handle and slowly chiseling away until you have something not only tractable but elegant.
I had a solid grasp of some of the key tools used in the process like user story mapping sessions, prototyping, jobs-to-be-done framing, sequencing, agile & waterfall and when to use each. Those tools are still the right vessels to think about the work but they don't necessarily look the same.
While it’s clear we are far from automating product discovery, it’s also a process where it’s difficult to escape LLMs. Product, Engineering and Design should still be involved but how do we use the 4th team member – AI? A large part of answer seems to reside in figuring out what a good design doc looks like.
A PE-backed company recently shared 4 (!) different design documents with me tackling the same product opportunity. One was written by a sales team member, another by the CEO, one by the engineer and another by a board member. Well, actually to some degree, all these docs were written by AI. They were also written from different and somewhat irreconcilable lenses.
While more of the plan was written down, the clarity and alignment wasn't there. AI should help us unify different perspectives not defend our individual forts.
If you think something will take a long time to build, AI will write a compelling document explaining why. If you think something should only take 3 weeks, AI will take the required shortcuts to get there.
It doesn't provide any judgment.
We've also been working on a very complex design document for a client and I can feel the opposing tensions of AI improving our work as well as adding friction, especially mentally.
It helps us understand third-party tech quicker but it also makes us feel stupid if we don't get it immediately from spending 10 minutes looking at the docs.
There’s a sense that things should move more quickly (because we don't have to do all the writing and prototyping?) but also there is no shortcut around the deep work.
We get more quickly to the initial version of a design document but it is not written and optimized from a place of clarity. So it doesn't compound our ability to close out the long tail of questions.
Because AI can help write the document, they are much bigger and complicated than if they were written entirely by a human.
We need design documents more than ever (to serve as a prompt) but that doesn't mean that is all we need them for. We also need them for clarity and consensus.
I can see that a lot of people have a very loose relationship with design documents seeing them merely as an iterative milestone. They will ask Claude to put one together, jump into implementation, jump back, and so on. I wish that worked for smart contracts but experience shows otherwise. Experience has showed me that it’s better to fear even one question left open, one feature missing or one edge case unresolved.
So I now value three things when working and designing with AI:
Simplicity. LLM’s natural instincts are verbose and entropic and setting a target on succinctness is a good way to ensure quality human participation.
Narrative. AI cannot do narrative, but narrative is what helps a document to be digested and reviewed meaningfully by other humans. It's essential to a document being useful at all beyond being fed into the next LLM pipeline step (in which case why surface the document at all?).
Reasoning. AI often presents the answer but without the irrefutable chain of logic that helps us “own” and defend it or even understand where it breaks down. A document edited and curated by a human can leave clear explanations of why we arrived at certain solutions at what else was explored.
Don't hesitate to sink more time in design docs and using your judgment; I promise you it will pay off.