The argument in one line.
The most valuable Claude Code plugins of mid-2026 all solve the same root problem: misalignment between what you want and what the agent builds, or between what the agent remembers across sessions.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code daily and feel like most plugin recommendations recycle the same five tools.
- You want a second AI model reviewing Claude plans before any code is written.
- You work with large codebases and notice token costs climbing as Claude greps through files repeatedly.
- You have an Obsidian vault and want Claude to automatically maintain and query it.
- You build frontends with Claude and the default output looks generic.
SKIP IF…
- You are completely new to Claude Code and have not yet internalized basic plan mode and CLAUDE.md usage.
- You want original research rather than a curated walkthrough of existing repos.
The full version, fast.
The core failure in AI-assisted development is misalignment -- the agent does not know what you actually want, and you do not know what the agent actually built. This video addresses that from three angles: better pre-build interrogation (grill-me, grill-me-codex), better cross-model review (Codex plugin), and better persistent memory (Graphify, claude-obsidian, notebooklm-py). The remaining tools -- Impeccable, Higgsfield, n8n MCP -- extend what Claude Code can produce rather than correcting how it plans. The most actionable single change is dropping a Karpathy-style CLAUDE.md with four principles into every project.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:34
01 · Intro
Hook on plugin noise, promise of 10 diamonds in the rough, none of the obvious big names.
00:34 – 02:13
02 · Graphify
Knowledge graph generation from codebases. AST-only rebuild (no LLM cost). Auto-rebuild hook on every commit. Obsidian vault export option.
02:13 – 03:19
03 · grill-me and grill-with-docs
Matt Pocock skills that go deeper than Plan Mode. Lightweight token cost. Solves alignment before code is written.
03:19 – 04:41
04 · grill-me-codex
Host-built extension adding Codex adversarial review on top of grill-me. Up to 5 rounds between Claude and Codex until both approve.
04:41 – 05:58
05 · Codex Plugin for Claude Code
OpenAI official plugin: /codex review, /codex adversarial-review, /codex rescue. Free tier available. Underused.
05:58 – 07:17
06 · claude-obsidian
Self-organizing Obsidian vault. Extracts entities from sources, updates hot cache at session end. No recap next session.
07:17 – 08:16
07 · Karpathy CLAUDE.md
Single CLAUDE.md with four principles. 170K GitHub stars.
08:16 – 09:38
08 · Impeccable
23-command frontend design skill. Live mode for real-time browser element editing. Before/after docs for all 23 commands.
09:38 – 10:15
09 · Higgsfield CLI and MCP
One-stop shop for AI image and video generators. CLI for automation, MCP for frontend design work.
10:15 – 11:19
10 · notebooklm-py
Unofficial Python CLI for NotebookLM. Batch downloads, PPTX slide decks, slide revisions. Free via Google.
11:19 – 12:07
11 · n8n MCP
Official n8n MCP server. Build workflows from Claude terminal. Self-hosted = free.
12:07 – 12:27
12 · Outro
Chase AI+ masterclass pitch.
Lines worth screenshotting.
- The most common failure in AI-assisted coding is misalignment -- not bugs, not hallucinations, but the agent not understanding what you wanted before it wrote a single line.
- Graphify creates knowledge graphs from codebases using AST only, so the LLM is never involved in the indexing step and rebuilds cost nothing.
- Plan Mode asks three questions; grill-me keeps asking until you and Claude are actually on the same page.
- A second AI model reviewing the plan catches structural problems Claude cannot catch about its own output.
- OpenAI Codex plugin for Claude Code has a free tier and has been live for months -- almost nobody is using it.
- claude-obsidian updates a hot cache at the end of every session so the next session starts with full context, no recap needed.
- The Karpathy CLAUDE.md has 170K GitHub stars for four principles that seem obvious until you realize Claude violates them constantly without them.
- Impeccable has 23 commands and a website showing before/after for each -- the live mode lets you point at browser elements and edit them in real time.
- notebooklm-py offloads AI tasks to Google servers at zero cost, and gives CLI access to features the NotebookLM web UI never exposed.
- The n8n MCP server works on self-hosted n8n, making it effectively free, and lets Claude build and update workflows without you touching the canvas.
Ten tools that close the gap between intent and output.
WHAT TO LEARN
The biggest failure in Claude Code workflows is not a bug in the tools -- it is misalignment between what you asked for and what the agent understood, compounded by the agent inability to audit its own plan.
- Graphify replaces repeated grep calls with a persistent knowledge graph built from AST analysis alone -- no LLM tokens consumed on rebuild, and Claude navigates large codebases more accurately as a result.
- Plan Mode stops at a few questions; grill-me and grill-with-docs keep asking until the agent and the developer are genuinely aligned, not just apparently aligned.
- A second AI model reviewing the plan before code starts catches structural problems the first model cannot see in its own output -- grill-me-codex and the Codex plugin both provide this at low or zero cost.
- Claude cannot reliably evaluate whether its own plan is the best path forward; Anthropic has acknowledged this, and cross-model review is the practical workaround.
- Four principles codified in a CLAUDE.md -- Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution -- constrain behavior across every project without requiring per-session reminders.
- claude-obsidian and notebooklm-py both solve the session-memory problem from different angles: one keeps a self-updating Obsidian vault, the other offloads processing to Google infrastructure for free.
- Impeccable live mode lets you iterate on UI by pointing at browser elements rather than writing prompts -- tightening the feedback loop between design intent and rendered output.
- The n8n MCP server works on self-hosted instances, meaning workflow automation from inside Claude Code costs nothing beyond the infrastructure already running.
Terms worth knowing.
- Knowledge graph
- A structured map of how every file, function, and dependency in a codebase connects to every other, enabling targeted retrieval instead of broad file search.
- AST (Abstract Syntax Tree)
- A parsed, language-aware representation of source code that captures structure without executing anything, allowing analysis at zero LLM cost.
- Plan Mode
- A Claude Code mode where the agent asks clarifying questions and proposes a plan before writing any code, reducing costly mid-build course corrections.
- Adversarial review
- A workflow where a second AI model evaluates the first model plan or output specifically to find problems the first model might miss about its own work.
- Hot cache
- In claude-obsidian, a lightweight summary file updated at the end of each session so the next session starts with recent context already loaded.
- MCP (Model Context Protocol)
- A standard for connecting Claude to external tools and data sources so Claude can call them directly from the terminal.
Things they pointed at.
Lines you could clip.
00:12
“It has gotten to the point where there is so much noise out there that most of us have stopped listening entirely.”
02:52
“The most common failure in software development is misalignment. You think the dev knows what you want. You then see what they built and you realize it did not understand you at all.”
03:49
“Claude Code cannot be trusted all the time to necessarily evaluate its own code correctly. This is something Anthropic has even brought up.”
07:41
“Instead of having Claude try to go as fast as possible, these guidelines bias towards caution.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Every day you're being told there's a new Cloud Code plugin that is going to change everything, and it's gotten to the point where there's so much noise out there that most of us have stopped listening entirely. But luckily for you, I've found some diamonds in the rough, 10 of them, in fact, and I am gonna be sharing them all with you today.
00:16These are 10 Cloud Code plugin skills and CLIs that will actually make a difference on your next project. And no, none of them are the big names you already know about, like superpowers or front end design.
00:26Instead, these are all relatively new. So unless you are as obsessed with this stuff like I am, I promise you will find something you haven't seen before. So the first one on our list is Graphify, and I actually did a full video on this yesterday.
00:39I will link that above. Now Graphify is all about being able to point Claude code at some sort of repo, some sort of code base, and create a knowledge graph out of it. When I talk about a knowledge graph, I'm talking about something like this.
00:52This looks super cool. Right? Visually stunning.
00:55But what it really is for Claude code is a map. It is . By creating this sort of knowledge graph with Graphify, we are able to have Claude code answer questions about a code base more effectively than without the graph.
01:11Normally, it uses something called the grep, but if we have the knowledge graph, we're able to answer the same questions with less tokens being consumed. Now Graphify comes with a ton of commands.
01:22But luckily, when you install this thing, it includes a Graphify skill, which teaches Claude code what command it needs to use for what job. And there are some interesting ones in here.
01:31If you just do forward slash Graphify, that's gonna have it run through the code base and create that knowledge graph you saw. But we can do things like add an Obsidian flag, and it will actually generate a brand new Obsidian vault for that repo you've just created. So if you're someone who loves Obsidian, but you wanna add a little more power to the memory layer, graphify makes total sense.
01:49Now for those of you who are looking at this knowledge graph, you're like, hey. Looks neat. But what happens when I update the repo?
01:54It's a living, breathing thing. Well, Graphify thought about that. You can run this hook command, which will auto rebuild after every commit so it stays updated.
02:03An auto rebuild sounds scary like it's gonna cost a bunch of APIs. That's not the case. It does AST only, which basically means the LLM isn't even involved.
02:11It's a deterministic thing. Now the second skill on the list is actually a couple skills, and it comes from Matt Pocock, and that is grill me and grill with docs.
02:21These are two skills that are essentially plan mode on steroids. I love plan mode. You should be using it before every single feature or project, but grill me and grill with docs, I would argue, are a little bit better because it does what plan mode does.
02:33It asks you a ton of questions so you and Clodco are on the same page, but it goes a bit deeper than the standard plan mode. And on top of that, these skills are pretty lightweight, so you aren't adding some huge token cost every time you spin it up.
02:46Now the big sell for Grille Me and Grille With Docs and why these skills exist is we are trying to solve the problem of the agent not doing what we want. As Matt writes here, the most common failure in software development is misalignment.
02:59You think the dev knows what you want, You then see what they've built, in this case, Cloud Code, and you realize it didn't understand you at all. And if you've just been using plan mode and you expect you and Cloud Code to be exactly aligned on the same page after, like, three questions, you're gonna be disappointed, and you've probably been disappointed in the past.
03:15And this is what Grille Me and Grille With Docs solves for. And piggybacking off of that is skill number three and that's Grille Me Codex. This is actually one I created.
03:24I did a video on this as well. And this takes Matt Pocock's Grille Me and Grille With Docs and adds a second layer to it. Because personally, I don't think it's enough for us to just get on the same page with Claude Code in terms of our plan.
03:39I wanna have a second set of eyes actually take a look at the plan Claude Code has come up with. Yeah. You can ask me all the questions you want, but do you actually understand what Claude Code built?
03:49Is it actually the best path forward? If you aren't an expert software engineer, you actually can't answer that question, and neither can Claude Code because Claude code cannot be trusted all the time to necessarily evaluate its own code correctly. This is something Anthropic has even brought up.
04:05So what this skill does is it brings in codex to review the plan in a read only sandbox. So you and Claude Code have a back and forth.
04:15It creates the plan, then Claude Code comes and takes a look. But it not only takes a look once, it has multiple rounds up to five rounds where Claude code and codex go back and forth. And at the end, it's gonna give you a thumbs up when they're both aligned.
04:29So we kinda get the best of both worlds. We get the best of Matt Pocock's grill me, and then we add a sort of codex adversarial review so you can be very confident that what you've come up with is the best path forward. And speaking of codex, that brings us into plug in number four, which is the official codex plug in for Claude Code.
04:46What I just described in the last one, you know, where we have this adversarial review and this back and forth, might be a bit much for some people. We don't need all that. We just want codecs to be able to come in and just do one pass at what we're looking at or attack one feature.
05:00And the codecs plugin is perfect for that. This is from OpenAI themselves, so this isn't just some random guy putting stuff together.
05:07This is from, you know, the actual creators of Codex. And when you install this plugin, you're able to use it inside of the Cloud Code terminal, and it can do things like review. Just do a basic review of what you've done.
05:18It has an explicit adversarial review, which goes deeper into some specific domains that you might not have thought of. And you can do things like Codex Rescue.
05:26So if you were building some specific feature and you wanted Codex just to work on that on its own, almost on the side, you can do that as well. And you can do this even if you aren't paying for OpenAI. There is a free tier for Codex now.
05:40Usage, you know, will be sort of an issue, but this is sort of like a great stepping stone. If you don't wanna go all in on Codex yet, but you want a second pair of eyes, but you don't wanna do the whole free local thing.
05:51So highly suggest checking this out. It's been out for a couple months now, but to be honest, it's hard to find people who are really using this in their day to day workflows. Now tool number five is Claude Obsidian.
06:01We all know about Obsidian. Right? It's a free tool that allows us to organize our markdown files, and it really came into vogue in the AI space when Karpathy came out and talked about his Wiki sort of style organization.
06:14Well, what if we had a plugin that organized it for us? That's exactly what this thing does.
06:19So this repo is meant to do everything Karpathy talks about when it comes to setting up your Obsidian Vault, but to automate the process. As it says here, you drop in the sources, whatever those are. Claude reads them, extracts entities and concepts, updates cross references, and files everything into a structured Obsidian vault and the Wiki or the vault itself continues to get richer and richer with with each ingestion, which means we can more easily or we can more effectively have Claude code take a look at a large corpus of documents, whether they're structured or unstructured, and give us correct answers.
06:54And what's also cool about this thing is the end of every session, Claude updates a hot cache. The next session starts with a full recent context. No recap needed.
07:03So if you're someone who's really interested in Claude and Obsidian and want sort of like an easy way to do all that for you and have all the organization be built in, highly suggest you check out this repo. Again, it is the Claude Obsidian repo.
07:17And speaking of Carpathi, we have to mention this one. This repo has a 170 k stars. It's a little bit older, but it's still one I wanted to bring up.
07:25And this is Carpathi's Claude dot n d file, and this is a great example of where we can make really effective changes with how we use Claude code in a rather simple manner because this is all it is. It's just a Claude dot MD file you can use, and it just talks about a few conventions that Claude should always follow.
07:42And they seem obvious on the surface, but there's also a reason this is so popular. The first one being think before coding.
07:50Number two, simplicity first. Three, surgical changes. Four, goal driven execution.
07:57And that's pretty much it. Again, it seems obvious and you almost wanna laugh like, okay. Yeah.
08:01Duh. You should do this. But there's something to be said about codifying this in a claw dot m d file.
08:07That way, this is always followed for everything you do. And instead of having claw try to go as fast as possible, these guidelines bias towards caution. Then moving on to tool number seven, we have impeccable, which is my favorite front end design tool in the game today.
08:22It's just one skill, but it encompasses 23 commands, and it's all about defeating the AI slop monster. Now the amount of commands it has at its disposal are quite large, like all these different things, colorize, animate, onboard, distill, quieter.
08:38How are you supposed to know what these things are supposed to do? Well, obviously, Claude code knows, but it helps for us to have an understanding of what's actually going on under the hood. And luckily, Impeccable makes it really easy because they have an entire website that breaks down with visuals what this actually brings to the table.
08:55For example, if I was wondering what Boulder does, and I will move this over here so it's easier to see, I can see what it looks like normally inside a Claude code, and I can see what it looks like with impeccable. If I was looking at animate, again, we have Claude code design and then we have impeccable's design.
09:13And we have this for 23 different sets, which is really, really cool. Another cool feature is that you can iterate on your UI with the live mode. So it'll start up a dev server.
09:23It'll actually bring up your website, and you can actually, like, point at different elements of your website and edit them with impeccable in somewhat of a live manner, which is really great because I know for me, I like to see the visual changes myself on the screen instead of always working in code. The number eight on the list is the Higgs Field CLI and Higgs Field MCP.
09:41I love this thing because I have a one stop shop for every single AI image generator, AI video generator. And this is something I use multiple times a week.
09:51I do a lot of content obviously, and a big form of content for me is things like carousels. And I'm able to create a full automation workflow which brings in certain AI image generators with the Higgs field CLI and just streamlines the whole process.
10:07Beyond that, things like Higgs field MCP are also really great if you're doing front end design work and wanna bring in image or videos into your UI. Now number nine on the list is Notebook LMPI, and I will never stop putting this in every single Claude code skill plugin CLI list I ever create because this is my favorite one.
10:24I use this literally every single day. This CLI lets you hook up Claude code to Notebook LM, and you get everything inside of Notebook LM and more with this tool.
10:34Why is this so good? Well, because NotebookLM is actually amazing, and it's free. You can offload tons of tasks to NotebookLM and Google servers, actual AI tasks that would otherwise cost you usage and tokens, and instead, NotebookLM does all of it for you.
10:49And like I said, the CLI actually gives you stuff that you can't do inside of the NotebookLM interface. Things like batch download, slide revisions, slide decks as PowerPoint, on and on and on. So I'm getting Notebook Element more and because it's a CLI, I can now integrate it into a number of skills.
11:06So I do tons of research related to YouTube that bring in Notebook LMPie because Notebook LM itself is really good at processing and dealing with videos, especially YouTube stuff because it's all under the Google umbrella. And last but not least, we have the official n eight n MCP server. N eight n still has a place in your AI toolkit.
11:25It isn't the dominant no code platform it was like a year ago, but if you're someone especially someone who works with clients and needs to create rather simple automations or workflows that the client themselves need to get hands on with, there's no better way to do that than n eight n. And the n eight n MCP is the best thing on the market.
11:44We had really hacky options in the past, but with this tool, which you can use on the self hosted version as well, by the way, so for all intents and purposes, it's free, you don't really ever have to even go into the n eight n canvas except to, like, do a final check.
11:59This works really really well. It's better than anything we've had in the past. So if you're someone who uses NA to N in any capacity, you need to use this MCP.
12:07So those are the 10 Claude code plugin skills and CLIs that I've been using in my projects lately and I think they can definitely improve yours. So as always, let me know what you thought.
12:18Make sure to check out Chase AI plus down in the pinned comment if you wanna get your hands on my CLOD code master class. And besides that, I'll see you around.
The Hook
The bait, then the rug-pull.
The Claude Code plugin ecosystem generates more noise than signal. A new tool drops every day with promises of workflow transformation -- and the fatigue is real. This breakdown cuts to the 10 that actually changed how the work gets done.
Named ideas worth stealing.
07:17list
Karpathy Four Claude Code Principles
- Think Before Coding
- Simplicity First
- Surgical Changes
- Goal-Driven Execution
Four CLAUDE.md directives derived from Andrej Karpathy observations on LLM coding mistakes. Bias toward caution over speed.
Steal forDrop into any project CLAUDE.md as a baseline behavioral layer
03:19model
grill-me-codex Two-Act Review
- Act 1: grill-me interrogates until plan is locked
- Act 2: Codex adversarially reviews plan in read-only sandbox
- Up to 5 rounds
- Thumbs up when both models align
A two-model pipeline that closes both the human-AI alignment gap and the AI self-evaluation gap before code is written.
Steal forHigh-stakes features or greenfield projects where a wrong plan costs hours
How they asked for the click.
VERBAL ASK
12:07product
“Make sure to check out Chase AI plus down in the pinned comment if you wanna get your hands on my Claude Code master class.”
Soft verbal CTA at end with pinned comment reference. No hard sell, no urgency language.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
