<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: DeadY2J</title>
    <description>The latest articles on DEV Community by DeadY2J (@deady2j).</description>
    <link>https://dev.to/deady2j</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4215577%2Fdb567b9d-4eff-4cd0-8dec-609a9f518a45.png</url>
      <title>DEV Community: DeadY2J</title>
      <link>https://dev.to/deady2j</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/deady2j"/>
    <language>en</language>
    <item>
      <title>I built a repo where a first PR is one folder, and a 144-line validator reviews the format</title>
      <dc:creator>DeadY2J</dc:creator>
      <pubDate>Tue, 03 Nov 2026 14:26:43 +0000</pubDate>
      <link>https://dev.to/deady2j/i-built-a-repo-where-a-first-pr-is-one-folder-and-a-144-line-validator-reviews-the-format-2a82</link>
      <guid>https://dev.to/deady2j/i-built-a-repo-where-a-first-pr-is-one-folder-and-a-144-line-validator-reviews-the-format-2a82</guid>
      <description>&lt;p&gt;I'm a beginner in open source. A few weeks ago a fairly big pull request of mine was merged into someone else's project, and the review taught me how much of a maintainer's time goes into checking things a machine could check.&lt;/p&gt;

&lt;p&gt;So I tried to design a small repo where that part is automatic: skills-es, an open catalog of Spanish-language skills for Claude Code.&lt;/p&gt;

&lt;p&gt;Disclosure: I built it with Claude's help and reviewed the result myself.&lt;/p&gt;

&lt;p&gt;The idea&lt;/p&gt;

&lt;p&gt;A "skill" is a folder with a SKILL.md file: reusable instructions an AI coding assistant can follow. Almost every catalog is in English, so this one is in Spanish.&lt;/p&gt;

&lt;p&gt;The contribution unit is one folder:&lt;/p&gt;

&lt;p&gt;skills/&lt;br&gt;
  my-skill/&lt;br&gt;
    SKILL.md&lt;br&gt;
    ejemplos/ejemplo.md&lt;/p&gt;

&lt;p&gt;Adding a skill never touches code. That matters for people sending their first PR.&lt;/p&gt;

&lt;p&gt;The validator&lt;/p&gt;

&lt;p&gt;Instead of the maintainer checking the format by hand, a Python script (standard library only, 144 lines) runs in CI on every PR. For each skill it checks:&lt;/p&gt;

&lt;p&gt;name is kebab-case and matches the folder name&lt;br&gt;
description is 20-1024 characters and has no &amp;lt; or &amp;gt;&lt;br&gt;
there is a "when to use" section&lt;br&gt;
an example file exists&lt;br&gt;
no secrets (ghp_..., AKIA..., private keys), binaries, or executables&lt;/p&gt;

&lt;p&gt;It returns all errors of a skill at once, not just the first, so a newcomer fixes everything in one round.&lt;/p&gt;

&lt;p&gt;Some choices I made:&lt;/p&gt;

&lt;p&gt;No YAML dependency. I parse the frontmatter by hand and only allow key: value on one line. It's less flexible, but there is nothing to install and the error messages are clearer. The trade-off: it will break the day someone wants a multi-line value.&lt;/p&gt;

&lt;p&gt;Tests that build skills in a temp folder. Each rule has its own unit test (13 in total) that creates a tiny skill, breaks exactly one thing, and asserts the error. Adding a rule means adding one test.&lt;/p&gt;

&lt;p&gt;Regex for secrets. Cheap and good enough for obvious leaks, but I'm not sure it's the right call versus a dedicated scanner in CI.&lt;/p&gt;

&lt;p&gt;Making the first PR easy&lt;br&gt;
CONTRIBUTING.md walks through fork, branch, folder, validate, PR.&lt;br&gt;
A PR template with a checklist.&lt;br&gt;
Ten issues labelled good first issue, each with acceptance criteria and a hint about where to look.&lt;br&gt;
What I don't know yet&lt;br&gt;
Whether a hand-rolled frontmatter parser is a mistake I'll regret.&lt;br&gt;
Whether secret detection by regex is enough.&lt;br&gt;
Whether anyone besides me needs these skills. That's the real test, and no amount of CI answers it.&lt;/p&gt;

&lt;p&gt;If you want to try it or send a skill, the repo is here: &lt;a href="https://github.com/DeadY2J/skills-es" rel="noopener noreferrer"&gt;https://github.com/DeadY2J/skills-es&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And if you've built contribution tooling for a small project, I'd love to hear what you'd do differently.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>beginners</category>
      <category>python</category>
    </item>
    <item>
      <title>How I stopped an LLM from inventing facts in my daily YouTube Shorts</title>
      <dc:creator>DeadY2J</dc:creator>
      <pubDate>Thu, 29 Oct 2026 15:09:50 +0000</pubDate>
      <link>https://dev.to/deady2j/how-i-stopped-an-llm-from-inventing-facts-in-my-daily-youtube-shorts-28cg</link>
      <guid>https://dev.to/deady2j/how-i-stopped-an-llm-from-inventing-facts-in-my-daily-youtube-shorts-28cg</guid>
      <description>&lt;p&gt;I run a channel that publishes one Short every day, unattended. The first version let the LLM write the script from a topic. It worked until it didn't: it slipped in numbers that weren't in my notes and presented rumours as facts. On a channel about facts, that is the one thing you can't afford.&lt;/p&gt;

&lt;p&gt;So I flipped the design. &lt;strong&gt;I write the verified facts once (with sources). The AI only turns them into a script, and the script has to pass guards written in plain code.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The guards
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Number whitelist.&lt;/strong&gt; Every number in the script must appear in the facts (plus a few allowed ones, like a countdown). Anything else is rejected and the model retries.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Repetition check.&lt;/strong&gt; Repeated n-grams and repeated ideas across recent videos are rejected.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rumour labelling.&lt;/strong&gt; Facts prefixed with &lt;code&gt;RUMOR:&lt;/code&gt; must be presented as rumours, or the script is rejected.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A second pass at temperature 0&lt;/strong&gt; that lists claims not supported by the facts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A facts-only fallback.&lt;/strong&gt; If the model keeps failing, the video is built straight from the facts. It is less punchy, but it is always correct.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What a reader taught me
&lt;/h2&gt;

&lt;p&gt;After I shared it, a reader pointed out the weak spot: guard #4 is another LLM call, so it is the only check that code doesn't verify, and the number whitelist can't see sentences without numbers (which are exactly the ones models invent most).&lt;/p&gt;

&lt;p&gt;His suggestion, which I'm tracking as &lt;a href="https://github.com/DeadY2J/factreel/issues/11" rel="noopener noreferrer"&gt;an issue&lt;/a&gt;: make the model cite, for every sentence, the index of the fact it comes from, and check in code that the sentence's numbers appear in &lt;em&gt;that&lt;/em&gt; fact. The index says where to look; the numeric check proves the sentence agrees with it. Each guard reinforces the other.&lt;/p&gt;

&lt;h2&gt;
  
  
  Fallbacks everywhere
&lt;/h2&gt;

&lt;p&gt;Same philosophy for the rest of the pipeline: voice (Edge TTS, then ElevenLabs, Cloudflare, offline system voice) and images (Cloudflare FLUX, NVIDIA, your own generator, archive, procedural gradient) each have a chain, so one service being down never stops the channel.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it / help
&lt;/h2&gt;

&lt;p&gt;It is open source (MIT, Python + FFmpeg), runs with zero API keys, and I built it with a lot of AI assistance, so design feedback is very welcome.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Repo: &lt;a href="https://github.com/DeadY2J/factreel" rel="noopener noreferrer"&gt;https://github.com/DeadY2J/factreel&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;&lt;code&gt;pip install factreel&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;Good first issues (new language packs, a JSON validator, local TTS): &lt;a href="https://github.com/DeadY2J/factreel/issues" rel="noopener noreferrer"&gt;https://github.com/DeadY2J/factreel/issues&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>python</category>
      <category>ai</category>
      <category>opensource</category>
      <category>ffmpeg</category>
    </item>
  </channel>
</rss>
