AI Development: Fast Waterfall vs. Software Dorodango Polishing

A recent essay by Jesse Vincent contrasts two AI development modes: a Superpowers/Claude/Codex-driven “fast waterfall” and a lightweight “Dorodango” polishing workflow.

AI Development: Fast Waterfall vs. Software Dorodango Polishing

TL;DR

  • Big up-front design (fast waterfall with agents): substantial planning and detailed specs (often thousands of lines); agents generate implementation plans and run full implementations; orchestrator runs end-to-end tests and captures screenshots/videos; one project succeeded on run 33.
  • Polishing workflow (software Dorodango): live product treated as a sculpting surface; small UI/streaming tweaks made with brief prompts and quick visual checks, avoiding full rebuilds.
  • Trade-offs: deep planning plus automated runs can produce complete implementations; polishing supports rapid, low-friction adjustments.
  • Agent iteration & artifacts: incremental runs produce partial artifacts that document gradual convergence and permit restarting from the original spec when intent diverges.
  • Framing: frequent agent edits reframed as deliberate, incremental craftsmanship rather than accidental mess.

A recent essay by Jesse Vincent, contrasts two AI-driven development modes: an extended up-front design path powered by Superpowers, Claude, and Codex, and a lightweight "polishing" path likened to the Japanese art of Dorodango.

Two development modalities

Big up-front design: a fast waterfall with agents

This workflow begins with substantial planning and specification. The planning phase can range from a few minutes to several hours of iterative brainstorming, often producing a detailed spec running into thousands of lines. From that spec, agents generate an implementation plan and then attempt a full implementation. The author notes use of an orchestrator that runs automated end-to-end tests and captures artifacts — screenshots and videos — as proof of progress. An illustrative detail: after many incremental attempts, a project finally succeeded on run 33, with earlier runs producing partial artifacts that tracked the agent’s gradual improvement. When results diverge from intent, the cycle often restarts from the original spec and a fresh implementation run.

Polishing workflow: software Dorodango

The second mode treats the live product as a sculpting surface for small changes. Typical work here is minimal, such as adjusting UI layout or tweaking streaming output. These changes are often accomplished with a brief prompt and a quick visual check, avoiding a full rebuild. The name Dorodango captures the essence: shaping and polishing a product until small rough edges smooth into a finished look. This daily refinement contrasts with the all-or-nothing nature of the big upfront approach and reframes frequent agent edits as deliberate craftsmanship rather than accidental mess.

Why it matters

  • Two distinct trade-offs are clear: deep planning plus automated runs can produce a complete implementation, while polishing supports rapid, low-friction tweaks.
  • The anecdote about incremental runs and captured artifacts highlights how agents can be framed as repeatable workers that converge on success through iteration.

Read the original essay for the full narrative and examples: https://blog.fsck.com/2026/02/10/dorodango/

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community