<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/4xQTDYdUH" -->

---
title: README Clew 2: Built with AdaL. Labeled by Jev. Judged...
description: A solo builder describes constructing README Clew 2, a tool that checks whether a public JavaScript/TypeScript GitHub README matches its code, built in five...
canonical: https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: README Clew 2: Built with AdaL. Labeled by Jev. Judged by Code. Explained by Haiku. | daily.dev
og:description: A solo builder describes constructing README Clew 2, a tool that checks whether a public JavaScript/TypeScript GitHub README matches its code, built in five...
og:url: https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh
og:image: https://api.daily.dev/og/posts/4xQTDYdUH.png
og:image:alt: README Clew 2: Built with AdaL. Labeled by Jev. Judged by Code. Explained by Haiku.
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# README Clew 2: Built with AdaL. Labeled by Jev. Judged by Code. Explained by Haiku.

**[DEV](https://daily.dev/sources/devto)** · 14 min read · 5 upvotes · 0 comments

## Summary

A solo builder describes constructing README Clew 2, a tool that checks whether a public JavaScript/TypeScript GitHub README matches its code, built in five days using AdaL (SylphAI's coding agent) for a hackathon centered on Jev, a small decision model from TypeSafe that labels claim types without generating prose. The architecture splits responsibilities: Jev labels what kind of claim a README line makes, five deterministic code checkers decide if it's true, and Claude Haiku writes a summary from counts only, never seeing the README itself to avoid prompt injection. The build log details test-driven development costs (474 tests, $40 agent bill), real bugs AdaL caught unprompted, and three false-positive bugs the tool itself exhibited when first run on the author's own old repo, all later fixed and turned into regression tests.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://dev.to/earlgreyhot1701d/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku-196e>

## Questions this post answers

### How can I prevent an LLM from being tricked by prompt injection hidden in a README file?

Keep the summarization model from ever reading the raw README text. In one implementation, a decision model labels what kind of claim each line is, deterministic code checks verify the claim against the actual codebase, and a separate model (Claude Haiku) writes the summary using only counts and names from the verified results, never the original README lines, so planted instructions in the README have no model to talk to.

_Developers hardening AI pipelines against prompt injection can track these architecture patterns on daily.dev._

### Why would a tool mark a correct README line as contradicted by the code?

False contradictions in README-verification tools often come from edge cases like monorepo workspace detection failing, package names appearing only in prose being misread as dependencies, environment variables referenced in files beyond a read limit, or README lines pointing at uncommitted build output like dist/. One tool fixed this by making unresolved cases fall back to 'unverifiable' rather than ever auto-labeling them as wrong.

_Anyone debugging false positives in doc-verification tooling can follow build writeups like this on daily.dev._

### What does Jev, the decision model from TypeSafe, actually do?

Jev is a small decision model that picks one label from a fixed set of options and returns a confidence score between 0 and 1, without generating any free text. It is accessed through a service called Glasser, which charges per call and caps each call at 100 questions. Its narrow job is classification, not explanation, in contrast to general-purpose LLMs.

_Developers evaluating narrow decision models versus general LLMs can keep up with real-world builds on daily.dev._

## Similar posts on daily.dev

- [Untitled](https://daily.dev/posts/untitled-nqahjlxwz) · SitePoint · 0 upvotes · 0 comments
- [Coding agents rediscover your codebase on every single run.](https://daily.dev/posts/coding-agents-rediscover-your-codebase-on-every-single-run--fvuu1qfcu) · Trendyol Tech · 4 upvotes · 1 comments
- [I paid people to try and follow my README](https://daily.dev/posts/i-paid-people-to-try-and-follow-my-readme-et5pin5vt) · Terence Eden · 1 upvotes · 0 comments
- [How We Use Claude Code at Serokell](https://daily.dev/posts/how-we-use-claude-code-at-serokell-xqpyzepya) · Serokell · 0 upvotes · 0 comments

---

Tags: [#ai](https://daily.dev/tags/ai), [#webdev](https://daily.dev/tags/webdev), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#testing](https://daily.dev/tags/testing), [#github](https://daily.dev/tags/github), [#typescript](https://daily.dev/tags/typescript), [#hackathon](https://daily.dev/tags/hackathon)

[View this post on daily.dev](https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"README Clew 2: Built with AdaL. Labeled by Jev. Judged by Code. Explained by Haiku.","url":"https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh"},"datePublished":"2026-10-13T04:25:47.336Z","dateModified":"2026-10-13T04:26:17.740Z","description":"A solo builder describes constructing README Clew 2, a tool that checks whether a public JavaScript/TypeScript GitHub README matches its code, built in five...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4abc84c1bf450f46a1e11684a1864edc?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4abc84c1bf450f46a1e11684a1864edc?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"DEV","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"DEV","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/devto","url":"https://daily.dev/sources/devto"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,webdev,llm,ai-agents,testing,github,typescript,hackathon","timeRequired":"PT14M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"DEV","item":"https://daily.dev/sources/devto"},{"@type":"ListItem","position":3,"name":"README Clew 2: Built with AdaL. Labeled by Jev. Judged by Code. Explained by Haiku."}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/readme-clew-2-built-with-adal-labeled-by-jev-judged-by-code-explained-by-haiku--4xqtdyduh#faq","mainEntity":[{"@type":"Question","name":"How can I prevent an LLM from being tricked by prompt injection hidden in a README file?","acceptedAnswer":{"@type":"Answer","text":"Keep the summarization model from ever reading the raw README text. In one implementation, a decision model labels what kind of claim each line is, deterministic code checks verify the claim against the actual codebase, and a separate model (Claude Haiku) writes the summary using only counts and names from the verified results, never the original README lines, so planted instructions in the README have no model to talk to. Developers hardening AI pipelines against prompt injection can track these architecture patterns on daily.dev."}},{"@type":"Question","name":"Why would a tool mark a correct README line as contradicted by the code?","acceptedAnswer":{"@type":"Answer","text":"False contradictions in README-verification tools often come from edge cases like monorepo workspace detection failing, package names appearing only in prose being misread as dependencies, environment variables referenced in files beyond a read limit, or README lines pointing at uncommitted build output like dist/. One tool fixed this by making unresolved cases fall back to 'unverifiable' rather than ever auto-labeling them as wrong. Anyone debugging false positives in doc-verification tooling can follow build writeups like this on daily.dev."}},{"@type":"Question","name":"What does Jev, the decision model from TypeSafe, actually do?","acceptedAnswer":{"@type":"Answer","text":"Jev is a small decision model that picks one label from a fixed set of options and returns a confidence score between 0 and 1, without generating any free text. It is accessed through a service called Glasser, which charges per call and caps each call at 100 questions. Its narrow job is classification, not explanation, in contrast to general-purpose LLMs. Developers evaluating narrow decision models versus general LLMs can keep up with real-world builds on daily.dev."}}]}
```

