Korva is a scroll of the internet’s best ideas.

Skipping the PRD

Roughly 60 to 80 percent of launches skip the PRD, and the remaining 20 to 30 percent get serious documentation because getting them right matters. Small changes, such as an upsell for something launching, are worked out in back-and-forth on Slack with a product-minded engineer who pushes back with questions about the audience.

The kickoff

Larger projects start with a cross-functional kickoff: legal, safeguards, and everyone else in a room for 30 minutes, with each person asked what they care about. That meeting streamlines the mess that would otherwise surface later.

When a PRD gets written

When a PRD does get written, it comes from a skill that holds the format and a project that holds all the previous PRDs, with the problem and the why fed in. Sometimes the prep is five minutes in Cowork before the meeting, producing a basic doc to talk from, and sometimes there is no doc at all.

Why did detailed requirements documents become standard?

Software was historically built in sequential stages, with requirements written first and fixed before design and coding began. Winston Royce's 1970 paper on managing large software systems described this sequence, which later became known as the waterfall model, though Royce himself warned that the pure sequence was risky. The Agile Manifesto of 2001 pushed against heavy paperwork by valuing working software over comprehensive documentation.

Why does an early kickoff meeting help?

A short meeting with legal, safeguards, and engineering in one room lets each group name its constraints before any design work exists. A concern raised at that stage changes a plan, where the same concern raised later forces a rebuild. Software engineer Barry Boehm's 1981 book Software Engineering Economics reported that the cost of fixing a requirements error rises steeply the later it is found. Later researchers have debated how steep that curve is in modern practice.

How does a skill hold a document format?

A skill is a folder of instructions and example files that Claude loads when a task matches it. Previous documents kept in a project serve as examples, and language models imitate the format of examples placed in front of them. The 2020 GPT-3 paper from OpenAI showed that a model could pick up a task from a few examples in its prompt, a behavior called few-shot learning.

More from Lenny's Podcast: Product | Growth | Career

Tech & AIClaude plays his boss each weekEvery week, Claude reviews a growth leader's work and gives feedback the way his manager would.
Enjoyed it?