4 Free Skills That Fix 95% of Claude Code's Problems

Claude Code out of the box is a fine coding agent. It is a weaker everything-else agent. Ask it to watch a 25-minute tutorial, sweep fifty YouTube videos for one topic, navigate a 10,000-file repo, or redesign a landing page without reinventing the same purple gradient — and it falls over in four predictable ways.

We pay for the top tier and use Claude Code daily, and we still hit those walls. None of us want to spin up a deep-research worker for what is a five-minute question or paste a YouTube transcript when the actual knowledge is on screen. So we did the obvious thing: paired Claude Code with four open-source skills, each one closing one specific gap. None of them costs a second subscription. All of them stay inside the terminal.

The four gaps Claude Code ships with

Every Claude Code session starts with a clean slate. Out of the box it cannot do any of these:

• Watch a video. Pasting a YouTube transcript is not the same as having the on-screen context — and the only major AI model that actually watches video today is Google's Gemini.
• Do quick research. WebSearch is fine for the surface, but above that you are one click from the deep-research worker, which burns ten million tokens on what could be a five-minute lookup.
• Remember a large codebase. Every new session re-reads everything you point it at. On a 10k-file repo that is a lot of tokens and a lot of waiting.
• Ship a UI that does not look like every other AI site. Inter for every font, a purple-to-blue gradient on every hero, cards nested inside cards, a rounded square icon tile above every heading — the defaults are obvious and boring.

One skill per gap. All four are open source, free, and live in the Claude Code plugin marketplace or on GitHub.

Claude Video — give Claude Code eyes

Claude Code cannot watch video. The transcript you paste is just the captions, and for a tutorial or a screen-recording the visuals carry half the meaning. The only major model that actually watches video today is Google Gemini, which means most people have been copy-pasting YouTube transcripts and hoping.

Claude Video, by Brad Bonanno (a.k.a. Brad Automates), plugs that gap with a single slash command. Pass a YouTube link or a local MP4 and the skill returns both a timestamped transcript and a sample of frames Claude can actually look at. Install via the Claude Code marketplace:

/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video

How it works under the hood:

• Captions first. Most YouTube videos already have captions, so that path costs nothing.
• Whisper fallback. For local files (Loom recordings, OBS exports), the skill transcribes with Whisper — Groq by default, OpenAI as a backup.
• Frame modes. Where it earns its keep. Twenty-four frames a second would get expensive fast, so the skill ships four modes that trade tokens for fidelity.

Mode	Frames	Best for
transcript	0	You only need the words.
efficient	Up to ~50 key frames	"What does this tutorial do?"
balance	Up to ~100, scene-aware	The default most people settle on.
token-burner	Uncapped (1,000+ on long videos)	Every detail, cost be damned.

Bonanno walks through the technical depth on his own YouTube channel if you want the under-the-hood version. Net effect for us: a capability Claude Code does not have natively, without paying Google for a separate Gemini API on top of what we already pay Anthropic.

NotebookLM-py — a free research terminal

Claude Code's built-in WebSearch is fine for surface-level questions. The problem is the gap between that and the deep-research worker, which costs an hour and a healthy slice of your token budget. We use notebooklm-py as the middle ground — an unofficial Python API and Claude Code skill by teng-lin that exposes everything in the NotebookLM web app from the terminal.

What it can ask NotebookLM to do:

• answer questions about a video (citations include timestamps, not blobs)
• generate slide decks, infographics and audio overviews
• build visual concept maps
• create custom documents, flashcards and quizzes

The catch is the same one NotebookLM itself has: it runs on Gemini, which is not as strong as Claude Opus or Fable for hard reasoning. The win is that inference is on Google's tab, not yours. The single biggest unlock for us: drop a stack of YouTube URLs on one topic into NotebookLM-py and ask it for a synthesis. Because the tool sits inside Google's ecosystem, it eats YouTube links for breakfast — pasting a long list of links is the actual workflow.

Installation is essentially "point Claude Code at the repo and let it pull the rest":

git clone https://github.com/teng-lin/notebooklm-py
pip install "notebooklm-py[browser]"

The [browser] extra installs Playwright the first time, which is the headless browser automation the skill uses to drive the NotebookLM web UI invisibly.

One caveat worth flagging: in July 2026 Google rebranded NotebookLM to Gemini Notebook. Same product, now also accessible inside the Gemini app. Old links redirect automatically, and the Python library keeps its original name.

Graphify — a knowledge graph of your repo

By "memory" here we mean how fast and how accurately Claude Code can answer questions about a large codebase or a pile of documents. Without help, every new session re-reads your repo from scratch. On a 200-file side project that is fine; on a 10k-file monorepo it is a tax you pay every morning.

Graphify, by Safi Shamsi, addresses the memory problem by building a knowledge graph of whatever folder you point it at. It parses code into nodes (via tree-sitter, locally — no LLM, nothing leaves the machine), then clusters nodes by content. The jargon means this in practice: when you ask Claude Code a question, it has a map to navigate instead of re-reading everything.

Graphify is not just for code. Point it at PDFs, screenshots, whiteboard photos or video and audio files and they fold into the same graph. That is the gap between "code search" and "I can ask anything about everything in this folder." Output ships to a graphify-out/ directory with an interactive graph.html, a GRAPH_REPORT.md plain-language overview and a queryable graph.json.

Where it earns its keep — and where it does not:

• Worth it on big repos, multi-folder research or anything where you keep asking the same scoping question.
• Overkill on a side project under a few hundred files — the built-in Read, Glob and Bash tools are faster.
• Bonus: ships an MCP server (graphify-mcp) so other tools can hit query_graph, get_node and shortest_path directly.

Impeccable — a design vocabulary the model obeys

Every AI model in 2026 trained on the same handful of SaaS templates, so every model defaults to the same tells:

• Inter for every font.
• A purple-to-blue gradient on every hero.
• Cards nested inside cards.
• Grey text on a coloured background.
• A rounded square icon tile above every heading.

The honest version: building a UI with AI in 2026 makes you look like a designer who has not designed anything. Impeccable, by Paul Bakaus, gives Claude Code a shared design vocabulary and 23 slash commands that catch the anti-patterns above before they ship. It started as an extension of Anthropic's own frontend-design skill and grew into a stricter, more opinionated tool.

Commands to give you the feel of how it works in practice:

• /impeccable init reads your tokens, components and Tailwind config, then writes a PRODUCT.md so every later command knows your brand and audience.
• /impeccable document captures an existing visual system into a DESIGN.md.
• /critique gives honest feedback on what you already built, including a numeric score.
• /colorize adds strategic colour to an interface that is currently all grey.
• /bolder and /quieter push a design toward louder or safer, depending on the diagnosis.
• /impeccable live brings the page up in a browser so you can click a component and iterate visually, instead of guessing at what changed from the terminal.

Impeccable also ships a detector with 61 deterministic rules — side-tab accent borders, purple/blue gradients, nested cards, low-contrast text, flat type hierarchy — that runs as a PostToolUse hook at edit time. No LLM, no API key, instant feedback when an anti-pattern lands in the diff.

What it does not do: it will not generate a design from a blank prompt, and it is not an accessibility audit. What it does well: catch the AI defaults before they ship, and give both you and the model a common language to fix them. In our testing on real landing pages, it sits one notch above Anthropic's bundled frontend-design skill and one notch above the UIUX-Pro-Max-style prompt packs that float around.

Install from the marketplace:

/plugin marketplace add pbakaus/impeccable

Or use the dedicated installer for tighter integration:

npx impeccable install --providers=claude --scope=project

Which to install first

None of these require a second subscription and none of them take you out of Claude Code. Video, research, memory and design — four gaps, four specific, free answers. Pick the one costing you time this week. The rest can wait.

If you genuinely do not know where to start, the answer is whichever task annoyed you most in the last seven days:

• Wasted an hour on a long tutorial? Claude Video.
• Spun up a deep-research agent for a five-minute question? NotebookLM-py.
• Re-explained your codebase to Claude for the fourth time today? Graphify.
• Shipped an AI-built landing page and grimaced? Impeccable.

They stack cleanly. Install all four, no conflicts, no exclusive registration — each skill only fires when you type its slash command. The wider set of free utilities for the rest of your stack lives in our tool directory; the developer-facing category is at developer tools; and the broader AI tooling roundup is at AI tools.

Frequently asked questions

Are any of these Claude Code skills built by Anthropic?

No. All four are third-party, open-source projects maintained by independent developers — bradautomates, teng-lin, Safi Shamsi and Paul Bakaus — and installed through Claude Code's plugin marketplace or straight from GitHub. They use Anthropic's skill API but are not officially part of Claude Code.

Do the four skills cost money on top of my Claude subscription?

No, none of them adds a subscription. Claude Video and Graphify run inside Claude Code, so their inference is billed the same way any other Claude Code session is — against your existing plan. NotebookLM-py routes through Google's Gemini tier, which has its own free quota. Impeccable is rules and slash commands with no extra API call.

Which skill should I install first?

Whichever gap you actually hit this week. If you do not know where to start, Impeccable gives the broadest day-to-day improvement because it touches every web project you open. Claude Video is the next-easiest win if you ever watch a tutorial or a Loom inside Claude Code.

Will these skills slow down Claude Code startup?

Each plugin loads at session start, so adding four of them makes the first turn after install a couple of seconds slower. After that there is no measurable hit — each skill only triggers when you call its slash command.

Do these skills work in Claude Desktop, or only Claude Code?

Claude Code only. The plugin marketplace, the /watch and /impeccable slash commands, and the /graphify pipeline are all Claude Code patterns. Claude Desktop does not load third-party skills from the marketplace today.