Chase AI argues that as good as Claude Code is out of the box, it still has clear weak spots, and that five free, open-source tools can patch them "drastically." He frames the video around five areas where Claude Code is "naturally sort of weakened": video, research, memory, front-end design, and token output. For each one he names a specific repo, explains the problem it solves, and shows how it plugs into your Claude Code stack. The recurring theme: you don't have to route through other paid APIs or build janky workarounds when a purpose-built skill or CLI already exists, and every tool here is free.
The first tool, Claude Video from Brad Automates (a little over 5,000 GitHub stars but trending hard), gives Claude the ability to ingest and watch video, a capability it lacks out of the box. Chase stresses this is not about generating AI video; it's about letting Claude actually see what is happening on screen. Today the only major model that natively does this is Gemini, so people who work with video are usually stuck with transcripts alone, and transcripts miss the on-screen context.
Claude Video is "the best of both worlds": it grabs the transcript and intelligently pulls screenshots/frames from the video when it deems them relevant. If no transcript exists (say, a Loom recording), it routes the audio through Grok's Whisper model (free) to generate one. Rather than throwing 24 frames per second at the model (which would get "crazy expensive"), it uses four modes to control how many frames are captured:
Installation is simple: add it from the marketplace or hand Claude Code the URL to the skill. Big picture, it gives Claude a brand-new capability without janky routing through Gemini or paying a Gemini API bill on top of Claude, keeping everything "in house."
The second tool targets research. Out of the box, Claude Code's simple web search is "fine but pretty surface level," and the only alternative is the heavy end of the spectrum, dynamic workflows and deep research that spin up "105 sub-agents and burn 10 million tokens." Most people want a middle ground. Enter notebooklm-py, which effectively brings NotebookLM inside Claude Code. It is both a skill and a CLI, functioning as an unofficial API into NotebookLM that can be called straight from the terminal, and it exposes everything the web version can do "and more."
The clever payoff: you are effectively getting free LLM calls. It's Gemini (not as powerful as Opus or Fable), but you can offload research and synthesis onto Google's servers for free, whether that's asking questions about videos or generating slide decks, infographics, or podcasts. Installation just needs the URL dropped into Claude Code; it will handle the rest, including setting up Playwright (invisible browser automation) if you've never used it. Chase's favorite use case is feeding it many YouTube URLs on a topic and synthesizing all of them at once, seamless because it lives under the Google umbrella (transcript-only for video, but very smooth).
The third area is memory: how to get Claude Code to quickly and effectively answer questions about very large codebases or large corpuses of documents. The tool is Graphify, which builds a knowledge graph of whatever codebase or document set you give it, breaking the content into nodes and clustering them by what they're actually about. The result is a map Claude can traverse, giving it a clear path from a question to an answer.
Crucially, Graphify is not a RAG system: no vector index, no embeddings, not LightRAG. Chase describes it as sitting "somewhere in between Obsidian and a true RAG system", a light version of graph RAG that is far less complicated than traditional RAG yet delivers many of the same memory benefits. It's also flexible on file types, handling not just Markdown but PDFs, images, video, and audio.
As a bonus tool, Chase highlights the Obsidian Skills repo, created by the CEO of Obsidian. It's a small handful of skills, but for anyone using Obsidian with Claude Code, it's an easy way to supercharge the workflow by teaching Claude Code best practices straight from the people who built Obsidian. "Don't sleep on this repo even though it's super simple."
The fourth tool, Impeccable, targets front-end design and is quickly becoming Chase's favorite design skill. It's not just heavily starred, it's officially part of GitHub's AI package. Impeccable is a single skill with 23 different commands, things like craft, shape, critique, layout, and colorize. For example, impeccable colorize adds strategic color to monochrome interfaces; the project website shows clear before/after comparisons of standard Claude Code output versus Impeccable output, and the Impeccable versions look noticeably nicer.
The standout feature is live mode, which "gives you shades of Claude Design." Running impeccable live brings your web page up on localhost in the browser, so instead of editing everything blindly through the terminal, you can click on individual components and see what they look like with and without Impeccable. It becomes a true visual design tool rather than a guess-and-check loop of "make that look nicer... okay, try again... make it more premium." Chase calls this "a huge step above the Anthropic front-end design skill" and above tools like UI/UX Pro Max, because you can see the change before you commit it.
The final tool, Ponytail, tackles token consumption, an ever-present concern, especially with an expensive model like Fable. Ponytail claims to make Claude Code 20% cheaper and 27% faster while producing the same results. It works by giving Claude Code a series of gates it must pass before writing code: Do you actually need to build this? Does the feature already exist? Is there a library for it? Only after passing does it get the "thumbs up" to build, and then only with the least amount of code necessary (the real logic is more complex than that, but that's the gist).
The benchmarks compare a gray baseline against green Ponytail runs, showing fewer lines of code, fewer tokens, lower cost, and less time. The catch: those benchmarks were run with a weaker model, whereas viewers are likely on Opus or Fable. Chase tested it himself, running Ponytail with Opus (even cheaper and quicker than the benchmark) and again with Fable, and across the board it reduced consumption while producing the same output. Benchmarks versus real life may differ by use case, but any chance to make Claude Code faster and cheaper at the same effectiveness is worth trying, worst case you do a couple runs and drop it. He also flags Caveman as a similar tool in the same vein worth a look.