Human Compiler
Web app that lints human prose like a compiler: Jev measures tone and information, code emits rustc-style diagnostics.
Overview
human-compiler is a joke made real: paste an email, Slack message, LinkedIn post, PR comment or memo and get compiler diagnostics in the style of rustc, with error codes, spans, notes and help. A code-only lexer finds phrases and counts features such as acronyms and emoji. Then one request to Jev asks a fixed set of typed questions about the text: about thirty yes/no Nouls, three Choices (intent, dialect, subtext), four Scores (information density, length excess, clarity, hostility), and one Noul per paragraph asking whether it adds anything.
Rules in code compare each measurement to a threshold and decide which diagnostics fire, so identical measurements always give identical output, which the test suite covers. Modes such as corporate, manager, linkedin, academic or reddit are lint profiles with their own thresholds; auto mode asks every mode's questions in parallel and picks the profile from the dialect answer, falling back to default below 0.50. Flags like -Wall, -Werror and -O2 change thresholds or strip paragraphs. It runs on Cloudflare Workers.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this paragraph add any information? | yes · no |
| Choice | Which dialect is this text written in? | default · corporate · manager · linkedin · academic · reddit · politician · teenager |
| Score | Information density, length excess, clarity, hostility | — |
How it works
- Lexer splits text and finds phrases with positions (no model)
- One Jev call answers ~45 Noul, Choice and Score questions
- Auto mode picks a lint profile from the dialect Choice
- Rules compare measurements to thresholds and emit diagnostics
- Render rustc-style report with severity and exit code
Strengths
- Model supplies measurements and code decides, so output is deterministic for given measurements
- All questions answered in one request (README shows about 812 ms for 45 questions)
- Thresholds and modes are explicit and adjustable
Caveats
- Primarily a humorous tool; scores are not validated writing metrics
- Requires a TypeSafe key to run your own instance
Getting started
Try the hosted version at https://human-compiler.asfarlab.fun.At a glance
- Type
- Demo
- Status
- beta
- Language
- TypeScript
- Runtime
- Cloudflare Workers, browser
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Paste corporate prose; Jev scores passive-aggression, urgency, and information density, then code emits rustc-style diagnostics. Live: [human-compiler.asfarlab.fun](https://human-compiler.asfarlab.fun).
AnotiaWang/awesome-decision-modelsA compiler for human language. Paste text, get diagnostics. Measured by TypeSafe Jev.
hellogumbo/awesome-jev
Frequently asked questions
What does human-compiler measure?
Things like passive aggression, urgency, information density and whether each paragraph adds anything, using Jev's typed answers.
Why is the output deterministic?
Jev only supplies measurements; fixed code rules decide which diagnostics fire based on thresholds.
What do the modes do?
Each mode is a lint profile with its own thresholds; auto picks one from Jev's dialect answer.
Related projects
Chrome and Firefox extension that scores a draft X post with Jev on a dozen rubrics and gives a send-or-don't verdict.
Browser extensionbetaJavaScriptChrome extension that labels the tone of Bluesky posts and drafts with Jev and can blur or collapse matching posts.
ServiceprototypeJavaScriptNext.js novel reader where Jev scores sentences near the viewport for eight emotions and highlights those above an adjustable threshold.
AppresearchTypeScriptAgent skill where Jev scores a task prompt's clarity, the agent refines it up to three times until it reaches 80/100, then builds.
ProjectbetaJavaScriptClaude Code plugin that logs your prompts locally and later has Jev score how well you prompt the coding agent over time.
PluginbetaTypeScriptResume review app and MCP server where a Cloudflare Worker asks Jev typed questions and composes a JevScore, generally or per job.
AppbetaTypeScript