<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Deep Learning on Top AI Skills</title><link>https://topaiskills.com/tags/deep-learning/</link><description>Recent content in Deep Learning on Top AI Skills</description><generator>Hugo</generator><language>en</language><lastBuildDate>Sat, 29 Aug 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://topaiskills.com/tags/deep-learning/index.xml" rel="self" type="application/rss+xml"/><item><title>I Ran repo-intake-and-plan on 3 Real Repos: What It Caught</title><link>https://topaiskills.com/tutorials/guides/repo-intake-and-plan-narrative-experience/</link><pubDate>Sat, 29 Aug 2026 00:00:00 +0000</pubDate><guid>https://topaiskills.com/tutorials/guides/repo-intake-and-plan-narrative-experience/</guid><description>&lt;p&gt;Repo intake sounds like a formality until the report lands. RigorPilot&amp;rsquo;s &lt;code&gt;repo-intake-and-plan&lt;/code&gt; helper (450,025 installs, #70 on skills.sh) is the scan-and-plan step of the family&amp;rsquo;s README-first reproduction pipeline, and today I ran its own scripts against three repos it would plausibly be handed: karpathy&amp;rsquo;s micrograd, makemore, and nanoGPT. Twenty-one commands came out of the three READMEs. One of them was a name the script invented, three more carried labels I&amp;rsquo;d have argued with, and the misses taught me more about the skill than the hits. Every number below is the scripts&amp;rsquo; actual output, run on shallow clones at 2026-08-29 17:06 UTC.&lt;/p&gt;</description></item><item><title>Reproduce an AI Paper Repo: The 9-Step Rigor Sequence</title><link>https://topaiskills.com/tutorials/guides/ai-research-reproduction-procedural/</link><pubDate>Fri, 28 Aug 2026 00:00:00 +0000</pubDate><guid>https://topaiskills.com/tutorials/guides/ai-research-reproduction-procedural/</guid><description>&lt;p&gt;&amp;ldquo;Run this repo&amp;rdquo; is the most dangerous sentence in deep learning. Pick the whole training script as your target and the run dies somewhere in epoch two, and you can&amp;rsquo;t tell whether the code is broken or your setup is. RigorPilot&amp;rsquo;s &lt;code&gt;ai-research-reproduction&lt;/code&gt; skill (310,832 installs, #162 on the skills.sh all-time board) exists to change that bet: it forces a README-first, smallest-honest-target pipeline that ends in a standardized &lt;code&gt;repro_outputs/&lt;/code&gt; evidence bundle. I read its SKILL.md, its patch policy, and its agent config on 2026-08-28. This guide walks the nine steps in order, with the exact rules that keep each one honest.&lt;/p&gt;</description></item><item><title>paper-context-resolver: 450K Installs, Zero Summaries</title><link>https://topaiskills.com/skills/general/paper-context-resolver/</link><pubDate>Thu, 27 Aug 2026 00:00:00 +0000</pubDate><guid>https://topaiskills.com/skills/general/paper-context-resolver/</guid><description>&lt;p&gt;paper-context-resolver exists for the moment a reproduction run stalls on paper details the README never explains. Your agent hits the evaluation-split section and starts guessing. That split could be any of three things. This 49-line helper from the RigorPilot family passed 450,765 installs on skills.sh by refusing almost every other job you could give it. I read its SKILL.md, its reference file, and the live leaderboard on 2026-08-27. The refusals surprised me more than the numbers.&lt;/p&gt;</description></item></channel></rss>