← All Posts

Productized SDD vs the Real Thing

Work

Part of the spec-driven development series

I spent an evening this summer reading a public methodology plugin for coding agents, obra/superpowers, skill file by skill file. It encodes, as rules the agent has to load, most of what I do by judgment. Reading someone else's rulebook for your own practice is a strange and useful exercise, because it shows you which of your habits are principles and which are just habits. This is what I found, as I read it in July; the project moves quickly and may have changed since.

What the Plugin Encodes

Plugin ruleWhat I doVerdict
A design gate before any code is writtenRoughly seventy percent of the effort goes in before the agent writes anythingAgree
Plans written for "an enthusiastic junior engineer with poor taste"The zero-edit PRD: complete enough that the agent needs no clarificationAgree, and the phrase is better than mine
A fresh subagent per task, with two-stage reviewScoped agents for bounded tasks; review is a human at the mergePartly
Test-first as an iron lawExample-based tests for the happy path, the edge, and the error case, written with the specPartly
A verification-before-completion skillThe agent will say it is done long before it is; taste the soupAgree

The two agreements at the top surprised me by how exactly they matched. The plugin's design gate is my seventy percent, written down as a rule the agent cannot skip. And "an enthusiastic junior engineer with poor taste" is the most precise description of the reader of a spec I have seen. It is the reader I write for. I had never named it.

What the Rigid Version Gets Right

A human improviser forgets the verification step on the day they are tired. The plugin does not. That is the entire argument for productizing a practice, and it is the same argument I made for gates: a preference decays, a rule keyed to an observable signal does not. For a team that does not yet have the instincts, a workflow that forces the design conversation, forces the plan, forces the test, and forces the verification is going to beat a team of improvisers most weeks, because the improvisers will skip one of those four on the weeks that matter.

It also makes the practice teachable. I have spent real effort building onboarding so that other product managers can stand up the same toolchain. A plugin that installs the method is a shortcut to the part of that I was doing by hand.

Where Lived Practice Diverges

Three places, and they share a shape. The plugin's two-stage review is agent reviewing agent. Mine ends with a person at the merge, and I have no plans to remove that person, because the checkpoint is where human judgment earns its keep. The plugin makes test-first a law, which is correct, but a law cannot make the tests good, and designing the evaluation is the hard skill. A rule that says "write tests first" is satisfied by bad tests. And the plugin has no notion of where the human's time goes. It treats the human as a reviewer at the end. My split is seventy up front, twenty reviewing, ten in the loop, and the up-front part is the job.

The fourth divergence is direction. The plugin brainstorms with you. My rule runs the other way: the agent interviews me, because I cannot see my own assumptions and its questions can. Ask, do not assume, is a rule about who holds the pen during the design conversation, and I want the agent asking.

The Honest Synthesis

A rigid workflow gets rightA human improviser gets right
It runs every step every timeIt knows which step matters today
It makes the method installableIt adapts the method to the domain
It cannot be talked out of verificationIt can tell good verification from theater
It scales to people without the instinctsIt builds the instincts

The productized version is a strong default for teams without the instincts, and I would recommend it to one without hesitation. The real thing is the same skeleton with a person placed exactly where judgment matters: at the evaluation, at the merge, and at the front of the process where the questions get asked. Read the plugin. Steal the phrasing. Keep the person.

0 Comments