stas4000

Jev Linkmap — AI skill for Claude Code

AI community

Rebuild a site's internal link map in seconds with Jev, race Claude Opus 5 on the same queue, and let a deep model rewrite the rubric from the disagreements.

How to install Jev Linkmap

This entry records only its repository, not the path inside it, so there is no exact command to give. Open stas4000/jev-linkmap and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Jev Linkmap does

Rebuild a site's internal link map in seconds with Jev, race Claude Opus 5 on the same queue, and let a deep model rewrite the rubric from the disagreements.

Alternatives in AI

Internal Safety Collapse — We built an adversarial codespace setup 1.2k ★
Clawcodex — Token efficient Claude Code full Python rebuild 875 ★
Obsidian + Claude Code: Rebuild Your Second Brain With AI — Build an AI-powered personal knowledge management system from scratch 540 ★

README

jev-linkmap

Rebuild a whole site's internal link map in seconds with [Jev](https://typesafe.ai). Run Jev alone (System 1), or put Claude or Codex behind it as System 2 to train its rubric on your site. Includes a race against Claude Opus 5 on the same queue.

Inspired by [borja's post](https://x.com/borjafat/status/2101018783976722479): internal linking is not writing, it is thousands of yes/no calls. Does this page have a real reason to link to that one, and is there anchor text already sitting in the copy. That is a classification problem, so it goes to a decision model, not a frontier writer.

[![The report](docs/report.png)](https://stas4000.github.io/jev-linkmap/)

Live report: https://stas4000.github.io/jev-linkmap/ · The race as a 22 second film: [video/jev-linkmap-race.mp4](video/jev-linkmap-race.mp4), built frame by frame from the real timings (`video/spec.cjs`).

The run in this repo

Site: www.bles-software.com, 566 pages, 8,460 link decisions (15 candidate targets per page). Run on 19 Sep 2026. Every number below is from the run files in `out/` and `runs/`.

	Jev (`typesafe/jev-1.13`)	Claude Opus 5
Pages finished when Jev was done	566	0
Wall clock	5.95 s, 32 requests at a time	first 8 pages landed by 16.5 s, 8 at a time
Cost per page	$0.00048	$0.118 (24 random pages, same rubric)
Whole site	$0.27	about $67
Links placed	679 on 334 pages
Pages left as they are	232, Jev links only when it is sure

Per page Jev is about 240 times cheaper. Crawling (17.6 s, 8 polite workers) and building the candidate queue (about 50 s of plain Python) are code, not model time, and are outside the race clock for both judges.

![The race](docs/race.png)

How it works

Code reads the site. crawl.py walks the sitemap and keeps title, headings, body copy without nav and footer, and the links each page already has. Standard library only.
Code does the search. candidates.py picks the