AI Software Company

Nine opinionated personalities that help move a project from roughly 80% done to a polished public release.

What It Does

I built this to help me get projects out the door. I tend to create many ideas to the point that they work and solve my problem but I am now interested in getting projects to a more polished state and this pipeline helps avoid the classic inventor dilemma of a) losing interest and b) kitchen sink.

How it works

Most of my personal projects hit the same wall: functional, roughly 80% done, then stalled because "polish" is a fuzzy word that hides five different jobs. What does done mean? What should be cut? Does it look designed or assembled? What breaks first? How would a stranger describe it? Is it actually ready to ship? Each of those questions wants a different reviewer with a different voice. Bundling them into a single "make this better" prompt produces flat, agreeable output. Splitting them into named specialists produces disagreement, and disagreement is where the sharp decisions live.

Personas, not pipelines. The specialists are invoked one at a time, in the main conversation, on artifacts you hand them. There is no orchestrator, no queue, no autonomous loop. The value is in the separation: five specialists looking at the same product from five angles will disagree (the Pruner wants to cut what the Designer wants to save), and that productive tension produces a sharper decision than any single reviewer could.

Every persona has a stated definition of failure. Not just success. The Pruner fails when it rubber-stamps existing scope. The Designer fails when it produces generic "clean up spacing" notes. Naming the failure mode inside the prompt makes drift toward blandness self-correcting: the model has to check its own output against a specific way it could fall short.

Explicit forbidden-actions list per persona. Alongside "responsibilities," each persona carries a list of things it must never do: no hedging, no filler encouragement, no breaking character, no suggesting features (Pruner), no discussing "as an AI." That list does more to preserve voice than the voice section itself.

Always ends with Top Three Actions. A shared closing structure across all six personas. No matter what a review contains, the reader can scroll to the bottom and see the three highest-ROI moves. The Product Clarifier is the one exception: as an intake rather than a review, it ends with Top Three Open Questions instead.

Built with

This is all Claude skills. I spent time exploring GStack, a similar concept made by Gary Tan but my pipeline is mostly advisory and discusses ideas, it does not implement directly without intervention.

The roster

Product Clarifier: the only interactive one; interviews you and writes SCOPE.md, the target the rest of the company argues against.
Product Pruner: reduces complexity until only the essential product remains. Every feature is a promise you have to keep forever.
Tasteful Designer: makes the software feel intentional instead of assembled. Reviews screens, hierarchy, typography, restraint.
Brutalist QA: breaks the software before users do. Ranks bugs by severity with repro steps, not hedges.
Launch Copywriter: writes store copy, onboarding, and release notes that make people immediately understand the product.
Release Manager: nothing ships accidentally. Runs a real go/no-go against the live build.

A sample persona snippet

Excerpt from personalities/product-pruner.md

Voice. Decisive. Unsentimental. You speak in short, load-bearing sentences. You are the person in the room who asks "why does this exist?" and waits for a real answer. You are not hostile to features; you are hostile to unjustified features.

Core beliefs. Every feature has a maintenance cost that outlives the excitement of building it. Hidden complexity is still complexity: a setting nobody sees still has to be tested. The best time to remove a feature is before it has users.

Forbidden. Never accept "it's already built" as a reason to keep something. Never let sunk cost drive a keep/cut decision. Never hedge: if it should go, say it should go.

Software-Company · personalities/product-pruner.md

How the sequence works

The personas can run in any order, but this one wastes the least work: you decide what ships before you polish it, break it, describe it, and gate it:

Clarify → cut → design → break → explain → gate.