Timothy Jordan

16 July

"The trouble is, the other side can do magic too." - Cornelius Fudge (This is a post about agentic coding)

Publishing an llms.txt is now nearly free, and adoption is climbing. Getting an agent to actually read it is the hard part. We benchmarked both sides. There's good news and bad news.

Not a just Claude thing. Ahrefs: 97% of llms.txt files got zero requests. Google: Search ignores them. The fix isn't "stop publishing," it's targeting: tell the agents YOU control to start from /llms.txt.

a14y.devResearch · a14yEmpirical findings about agent readability. How much does raising a site's a14y score actually save an AI agent? Measured results from running real coding agents against real websites at different scores.

Full benchmark on a14y.dev/research, and I'm sending these first on Substack: substack.com/@EngAndArt

Publishing an llms.txt is now nearly free, and adoption is climbing. Getting an agent to actually read it is the hard part. We benchmarked both sides. There's good news and bad news.

So how does the agent decide to use it? We served identical files and only changed how the page pointed at /llms.txt. Head tag: 0 fetches. A second alternate link: 0 fetches. A visible "For agents" footer link: 0 fetches.

Then we changed one line of the prompt ("check for /llms.txt first") and removed the link entirely. Adoption: 5 of 5. Tokens down about a third

Publishing an llms.txt is now nearly free, and adoption is climbing. Getting an agent to actually read it is the hard part. We benchmarked both sides. There's good news and bad news.

The payoff is real: a clean markdown discovery layer cut an agent's tokens roughly in half on the same task, same answer quality. Pulling a tight index beats crawling and re-parsing HTML.

I scored the 50,074 most-visited websites for how readable they are to AI agents. Not one scored "excellent." The median site is a 52 out of 100. Here is what the web looks like to an agent.

It is all open: open spec, open CLI, and a public leaderboard of all 50,074 sites. Check your own in a few seconds: npx a14y check yourdomain.com --mode site

The web has SEO-era plumbing but has not built the agent-era layer yet. The full reports and the leaderboard are live. I would love to hear what you find on your own sites. a14y.dev/research/

I scored the 50,074 most-visited websites for how readable they are to AI agents. Not one scored "excellent." The median site is a 52 out of 100. Here is what the web looks like to an agent.

This is fixable in a single coding session. I raised one site from 37 to 89 and Claude used about 49% fewer tokens to do the same task, with no drop in answer quality.

What actually pays off? We toggled 11 features on and off. A markdown mirror alone saves about 48% of tokens. A real meta description is irreplaceable. The agent-skills directory actually hurt. Ship the mirror and the description first.

I scored the 50,074 most-visited websites for how readable they are to AI agents. Not one scored "excellent." The median site is a 52 out of 100. Here is what the web looks like to an agent.

Agents (Claude, ChatGPT, Perplexity, Cursor) are already primary readers on the web. When one lands on your site, it has to find, read, and understand your pages fast. Most of the popular websites make it slow and expensive.

Only about 1 in 4 sites ship an llms.txt or an AGENTS md. 44% score under 50. The single best site in the top 50,000 managed an 83.

Everyone selling "AEO" slot content right now is reinventing the link farm. We've seen this movie before, when webmasters tried to cheat PageRank. We know how it ends.

Google already says ignore "inauthentic brand mentions" and that AI optimization is still just SEO. The durable play is the one SEO landed on: be agent-readable, then actually be the best. Everything else is borrowing visibility from a future penalty. Full piece below.

timothyjordan.com/blog/2026/0...

timothyjordan.comThe Ghost of PageRank: AI Spam Will BackfireIn the early days of web search, webmasters tried to cheat search algorithms. We’re seeing the same thing playing out by those who are trying to game model training with slop content.

Everyone selling "AEO" slot content right now is reinventing the link farm. We've seen this movie before, when webmasters tried to cheat PageRank. We know how it ends.

The old exploit: if a link is a vote, manufacture votes. Spin up pages that link back to you, rank climbs. Google answered with penalties for link schemes + a Manual Actions report. Link farms went from edge to liability.

Today's version: generate piles of "top 10 best tools" lists with your company on top, seed them where crawlers look, hope the model ingests it. It appears to work, only because the algorithms haven't caught up to penalize it yet.

youtube.comVercel Product Walkthrough (2026)Speed run a web app into modern infrastructure with AI features, agent-driven development, and collaboration between designers and developers.Timothy Jordan,...

23 Mar

In case you missed it, we updated the Vercel Product Walkthrough. www.youtube.com/watch?v=zFX...

4 Mar

This also works for getting your existing agents to run on Slack. You don’t have to rebuild from scratch. Same skill, just invoke with "/slack-agent config" instead of "/slack-agent new". x.com/vercel_dev/...

vercel.comAgent Skills: Creating, Installing, and Sharing Reusable Agent Context | Vercel Knowledge BaseThis guide will cover what skills are, how to create custom skills for yourself and your team, and how to publish them to the community.

17 Feb

Agent Skills let you package context, workflows, and domain knowledge into reusable building blocks for AI agents. This is how you scale AI without copy-pasting prompts everywhere. vercel.com/kb/guide/ag...

17 Feb

Claude Code pro-tip: adding "--scope user" at the end of a "claude mcp add" command will make that MCP have global scope. This isn't terribly well documented, but the obvious approach for something like the @vercel or @linear MCPs.

16 Feb

Claude Code pro-tip: every time you "always allow" a command, Claude adds it to that project's settings.json. You should regularly curate some of these into ~/.claude/settings.json

14 Feb

A lovely read to end your week: "When everything became available, taste became priceless." substack.com/home/post/p...

vercel.comDeveloper Advocacy DirectorJoin the team.

14 Feb

Come work on my team at Vercel. We're redefining developer experience. vercel.com/careers/dev...

13 Feb

Have you "npx skills update" today?

21 Jan

Game changer: npx skills add <owner/repo>

3 Nov 2025

I’ve redeployed timothyjordan.com to @vercel.com, and I’m super excited to lead the Developer Experience team at a company famous for excellent Developer Experience.

26 Oct 2025

My exit from Google: we don't say goodbye, we say good journey. timothyjordan.com/blog/2025/10...

25 Aug 2025

I just made Macaroni and Cheese and sliced hot dogs for my kids. Because apparently history does repeat itself.

20 May 2025

The Instagram backstage pass has started: www.instagram.com/stories/goog... #GoogleIO

20 May 2025

I've been to nearly every #GoogleIO and I'm still getting excited the night before. Can't wait to see old friends, help introduce some new things, and have great conversations about innovation and technology. goo.gle/IO25_DevKeyn... io.google www.instagram.com/googlefordevs/

18 Mar 2025

At GTC today, DM me if you wanna say hi.

blog.googleIntroducing Gemma 3: The most capable model you can run on a single GPU or TPUToday, we're introducing Gemma 3, our most capable, portable and responsible open model yet.

13 Mar 2025

Open Models, FTW.

Jeff Dean@jeffdean.bsky.social

Introducing our Gemma 3 open models, the most capable models that you can run on a single GPU or TPU. Multimodal, multilingual, 128k context length, and exceeds quality of other open models that are an order of magnitude larger in terms of hardware footprint. 🎉 blog.google/technology/d...

youtube.comHow AI is impacting the developer experience: DevRel insights from Timothy JordanYouTube video by All Things Open

22 Feb 2025

Caption this, mine: "I once caught a fish this big." I sat down with the ATO team to discuss Open Source, how generative AI will impact Developer Relations, and some best practices for using AI in your company. Video: www.youtube.com/watch?v=g602... Article: allthingsopen.org/articles/ai-...

io.googleGoogle I/O 2025Don't miss Google I/O, featuring product launches, innovations, and insights. Tune in for the live keynotes and sessions.

12 Feb 2025

I hope you see you at #GoogleIO on May 20-21, 2025. Live streamed keynotes, sessions and more: io.google

opensource.googleblog.comSee the code that powered the Pebble smartwatchesThe source code for the Pebble operating system in now open, to support the volunteer community maintaining Pebble watches.

28 Jan 2025

This is cool. We've open sourced the code that powered Pebble smartwatches. Check it out: opensource.googleblog.com/2025/01/see-...

news.ycombinator.comWe're bringing Pebble back | Hacker News

28 Jan 2025

Also, this: news.ycombinator.com/item?id=4284...

puf.ioAstro SSG: handling redirects on GitHub PagesI use Astro to generate a fully static site from content in Markdown files, that I then host on GitHub Pages. I want to provide incoming links/aliases inside my Markdown files, but GitHub Pages doesn't support custom 301 redirect responses. Read the full article to learn how I work around this limitation.

23 Jan 2025

I'm a big fan of static site generation. I use Jekyll with Firebase Hosting (it can also be used with GitHub), but same paradigm. This post makes me want to check out Astro as well, @puf.io !

Frank van Puffelen@puf.io

I use Astro for puf.io. From markdown files it generates a static site that I host on GitHub Pages. Sometimes I include inbound redirects (e.g. old URLs) in the .md, but Pages doesn't support custom HTTP status codes like 301. Solution: puf.io/posts/astro-s...