Show HN: Huzzah – a novel approach to coding with AI | AI Paper Digest

TL;DR Highlight

Save a pseudocode file and its diff becomes the prompt that makes AI regenerate code

Who Should Read

A developer tired of writing long English prompts to coding agents and frustrated that no record of intent survives

Core Mechanics

Agents leave no record of intent
- The author says prompts are discarded and it's unclear whether code was AI-generated.
- The central document expressing what the human wants from the machine is gone.
Chat prompts repeat the same instructions
- Chat describes step-by-step changes to the app, not the app itself.
- So instructions get repeated and tokens are spent many times over.
Huzzah makes prompts pseudocode, declarative, persistent
- Agent prompts are longform, imperative, transient; Huzzah prompts are the opposite.
- Files use the .hz extension.
Saving the file turns the diff into the prompt
- Saving a .hz file makes Huzzah generate real code automatically.
- On edits, the diff goes to the LLM and only affected source is regenerated.
Pseudocode doubles as documentation

Community Reactions

Pushback: it's just a new language
- One commenter said it's a terse language that costs money to compile.
- Another joked that Python is deterministic and needs no subscription.
Opinion that you can already do this
- Commenters said you can write pseudocode and ask Claude for real code today.
- One suggested a system prompt telling the agent to spell out intent, then implement and test.
Concern about large codebases
- A commenter said big codebases would need a pseudocode twin, which is extra work.
- Another noted complex intent makes pseudocode nearly as long as code.
View that fatigue comes from pace, not English
- One commenter argued agent work delegates thinking, so pseudocode won't ease the delegator's burden.
Declarative spec tool mentioned as alternative
- One commenter built spekk-cli to keep only intent and assertions that must hold.

How to Apply

Throw pseudocode at Claude
- Pick a small function today, write it as shorthand pseudocode, and ask for a real implementation.
- Paste the prompt below and check whether the output matches your intent.
Commit intent files with code
- Keep human-written pseudocode files per feature in the repo and include them in PRs.
- To change behavior, edit the file and use its diff as the prompt.
Try it on new modules only
- Avoid existing code or heavy cross-file dependencies and test on an independent new module.
- Note where output drifts from intent to learn the limits.

Try It

copy & run

# Paste into a Claude or ChatGPT chat (simplified version of the Huzzah approach)

너는 의사코드를 실제 코드로 바꾸는 도구다.
규칙: 1) 먼저 내 의도를 한 줄씩 풀어 쓴다. 2) {{언어}}로 구현한다. 3) 테스트도 함께 작성한다.

의사코드:
{{의사코드}}

Next to try: give a diff changing one pseudocode line and ask "regenerate only the changed parts."

Terminology

PseudocodeA design note that writes logic simply without following a specific language's syntax, like a recipe summary.

DeclarativeDescribing what the result should be rather than how to get there, like giving a destination address instead of directions.

ImperativeGiving step-by-step instructions on what to do, as chat prompts typically do.

HarnessThe runtime wrapping an LLM with tools and context, like a coding agent app.

Spec-driven developmentWriting a specification before code and implementing against it.

Show HN: Huzzah – a novel approach to coding with AI