57 days vs 818 days of estimated engineering work. Which would you trust?

[#1]

I ran a little experiment for the same feature with two different PRDs.

• PRD A: I fed ChatGPT some context and asked it to "write the PRD." → Estimated implementation time: 57 eng days.
• PRD B: I used Wisary in Confluence, which interrogated me with questions, forced me to slow down, and think more thoroughly → Estimated implementation time: 818 eng days.

At first glance, 818 days feels like "bad news." But in reality, it's the first time the full scope of work became visible.

What inflated the estimate?

• Security, reliability, performance
• Integrations, auth, permissions, audit logs
• Metrics, rollback, observability, supportability
• Testing, release strategy, change management

A lightweight PRD skips these decisions, making things look fast. Until launch time exposes the gaps.

👉 The real lesson: engineers don't underestimate because they're bad at estimating. They underestimate because the scope isn't visible.

Estimates are an output of clarity.

A rigorous PRD slows you down just enough to go faster, because once the work is visible, you can actually decompose, challenge, and right-size it.

Question for this group:

How do you balance the need for speed in writing requirements with the discipline needed to make hidden scope visible early?