Huzzah: an experimental editor that replaces chat prompts with pseudocode

Software engineer Daniel Vaughn has published Huzzah, an experimental editor that proposes an alternative paradigm for working with large language models. Writing in August 2026, he describes the fatigue of composing long English instructions for every change to a codebase, and argues that the durable part of a prompt — human intent — should live in a file rather than in a chat window.

The trouble with coding agents

Vaughn lists three problems. First, there is no reliable record of human intent: prompts are discarded, and the code may or may not have been generated by AI, so the central authority expressing what the human wants out of the machine is lost. Second, AI chats are imperative, step-by-step instructions that describe changes to an application rather than the application itself, which means the same guidance is repeated and consumes tokens many times over the course of development. Third, much of natural language exists for social reasons rather than informational ones; the average sentence is scarce in real information, which makes writing to a machine cumbersome.

How Huzzah works

With coding agents, prompts are longform, imperative and transient. With Huzzah they are pseudocode, declarative and persistent. A developer creates a file such as fizz_buzz.hz and writes a compact representation — a function, a loop over 100 iterations, checks for divisibility by three and by five. Saving that file makes Huzzah generate the real code. When the developer edits the file, the editor captures the diff and uses it as the prompt to the model, and the affected source is regenerated. The post demonstrates the same pattern for a shopping cart and a todo list.

Benefits, caveats and current state

The pseudocode is more terse and readable than longform prompts, engages the developer in designing the shape of the code, and doubles as developer documentation because a human wrote it to express intent. A language-agnostic description could serve as the basis for several language or environment targets, which the author suggests may suit complex algorithms such as a CRDT. He is equally clear about the limits: the approach may run into problems at scale, it fits new codebases better than existing ones, cross-file dependencies may be difficult to express reliably, LSP-style features would not be available, and without domain expertise natural language is probably the easier interaction method.

Huzzah exists only in an experimental state and is under active development. The source code and setup instructions are published on GitHub at github.com/danielvaughn/hz, and the author invites developers to try it and send feedback.