The argument in one line.
Most published Claude skills are junk, but nine specific skills across three categories transform how you use Claude, and the learnings.md pattern makes every skill compound over time by logging corrections into institutional memory.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude daily for repetitive workflows (writing, analysis, coding) and want to stop re-explaining the same instructions in every new chat.
- A builder or automation enthusiast who knows basic prompt engineering and wants to unlock Claude Code skills and system-level customization without deep programming knowledge.
- Someone managing Claude across multiple projects or team members who needs institutional memory and wants a concrete system to capture and reuse what actually works.
SKIP IF…
- You're new to Claude and haven't established a consistent workflow yet — this assumes you already know how you use Claude and what repeats.
- You're looking for advanced multi-agent systems or production-grade automation — this covers individual productivity skills, not orchestration at scale.
The full version, fast.
Most Claude users repeat the same prompts forever instead of building reusable skills that compound. The video filters 250+ published Claude skills down to nine worth keeping, split across three buckets: Claude Desktop essentials (Skill Creator to turn repeated workflows into skill.md files with side-by-side test cases, Prompt Master to restructure messy brain dumps, Fact Checker to verify claims before publishing, Humanizer to strip AI tells using your own writing samples), Claude Code skills (Playwright plus Superpowers for tested builds, a Septum-style sub-agent team, /review and /security-review, and /context plus /compact for session hygiene), and one co-work pattern. The pattern that matters most is a learnings.md file alongside any skill, logging corrections so the skill internalizes your preferences instead of repeating mistakes.
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 – 01:17
01 · Intro — the problem with Claude skills
Most people use Claude at its most basic. Skills can unlock catch-bugs, monitor, and behave-exactly modes — but 250+ skills tested, most are garbage. Three categories, one map.
01:17 – 01:33
02 · Category 1 intro — Essential Claude Skills
Work in Claude Desktop. No terminal, no code. Just upload a zip through Customize.
01:33 – 03:18
03 · Skill 1 — Skill Creator
Pre-installed. Describe a repeated workflow, Claude interviews you, builds the .md file, then auto-generates side-by-side test cases to verify improvement before you commit.
03:18 – 05:08
04 · Skill 2 — Prompt Master
GitHub: nidhin-js. Turns messy brain dumps into structured, optimized prompts. Detects target AI tool, extracts intent across 9 dimensions, max 3 clarifying questions. Works with 30+ tools. Pro tip: add to General Instructions for auto-enhancement on every complex prompt.
05:08 – 06:20
05 · Skill 3 — Fact Checker
GitHub: Daymaid's Claude Code Skills repo. Runs systematic fact-verification on any text — confirmed / unverifiable / false. Three use cases: public AI-generated content, own video outlines pre-film, other people's claims.
06:20 – 07:53
06 · Skill 4 — Humanizer
GitHub: Blader's humanizer repo. Detects 25 AI writing patterns (m-dash overuse, rule-of-three, leverage/streamline vocabulary, vague attributions, filler phrases). Feed it a writing sample and it mirrors your sentence rhythm instead of producing generic clean output.
07:53 – 09:34
07 · Category 2 intro + Skill 5 — Playwright + Superpowers
Claude Code skills operate at a different level. Playwright (Anthropic official) navigates your app like a real user — clicks, forms, screenshots at every step. Superpowers (Obera's repo) adds plan-before-code and test-before-ship quality gates baked into every build step.
09:34 – 11:35
08 · Skill 6 — Septim Agents Pack
Drops 10 named sub-agents into your project: Atlas (planning), Luca (architecture), Canon (brand), Amber (marketing), Tally (finance), Nova (design), Ward (legal), Meta (CX), Juno (research), PIP (coordination). Original costs $49; creator built a free version in Skool community.
11:35 – 12:21
09 · Skill 7 — /review + /security-review
Type /review after any build — Claude checks bugs, edge cases, security issues, design problems. /security-review shows a security audit of pending changes on the current branch. Runs locally, costs nothing extra.
12:21 – 13:16
10 · Skill 8 — /context + /compact
Every tool dumps raw data into the context window. /context shows what's eating it. /compact with a focus argument (e.g. 'compact focus on auth module') compresses on your terms. Habit: run /context every 20 minutes, /compact when you hit 60%.
13:16 – 16:11
11 · Category 3 + Skill 9 — learnings.md pattern
No install needed. Create learnings.md alongside any skill. Add one line to the skill: 'before running, read learnings.md'. After every run, Claude logs what worked, what failed, what you corrected — scoped to that skill. Skills compound instead of repeating the same mistakes.
Lines worth screenshotting.
- A learnings.md file alongside any skill converts a static set of instructions into institutional memory: every correction the human makes gets logged, and the next run reads those notes before executing.
- Installing a skill does not make you better at using Claude — a skill that you never improve is just clutter, and the learnings file is the only mechanism that makes improvement automatic rather than manual.
- Testing a skill with it active and without it, side by side, is the quality gate that separates skills that actually work from skills you assume are working based on a favorable first impression.
- Prompt master rewrites your messy brain dump into a structured, optimized prompt before Claude processes it — one extra step that produces dramatically better outputs on complex tasks, with no permanent cost to the workflow.
- A fact-checker skill that runs a systematic verification pass on AI-generated text before it goes public catches the specific category of error that destroys credibility: confident-sounding claims that are simply wrong.
- A humanizer skill that analyzes a sample of your own writing and applies your sentence rhythm, word choices, and quirks to AI output produces something that sounds like you rather than generic clean AI prose.
- Context window management as a skill — running slash context every 20 minutes and compacting on your own terms rather than letting Claude auto-compact at random — prevents the session degradation that makes AI sessions feel unreliable after an hour.
- The Playwright + Superpowers skill combo catches the category of bug that no amount of code review finds: a button that is technically present but visually hidden on screens under 400 pixels wide only appears in a real browser interaction test.
- A skill that plans architecture first, writes tests alongside code, and does a self-review before showing output has quality gates baked in — which changes the human's role from QA department to approver.
- Multi-agent setups in Claude Code where Atlas plans, Ward checks for legal issues, and Tally estimates costs in one project session give you three expert perspectives on a problem without re-explaining context to each one.
- Slash review as a final step before merging catches things Claude missed on the first pass — because a code-generation model and a code-review model are applying different heuristics to the same file.
- 250+ skills tested, 9 kept — the filtering ratio reveals that most published skills are broken, outdated, or renamed system prompts, and the correct investment is curation rather than collection.
- Skill creator as the first skill to use — because it interviews you about your workflow and builds the skill file from conversation rather than requiring you to write markdown — removes the technical barrier that prevents most users from building custom skills.
- Adding a fact-checker and a humanizer as automatic final steps inside other skills means you never manually invoke them — the quality gates run invisibly on every output without adding decision points to the workflow.
- The difference between a calculator and a computer is the learnings file: one executes the same operation forever, the other accumulates state and improves the result of every subsequent operation based on prior experience.
The one pattern that makes everything compound.
Steal this for your own skills
learnings.md is a free, one-file upgrade that turns any static skill into a system that gets smarter every time you use it.
- Add learnings.md alongside every skill in ~/.claude/skills/ — it costs two minutes and zero installs.
- Add one line at the top of each skill: 'Before running, read learnings.md and apply logged corrections.'
- After each run, have Claude append what it got wrong and what you corrected — scoped to that skill, not a general note.
- The Playwright + Superpowers combo is the most actionable for JoeFlow solo dev — it catches visual/responsive bugs that only show up when a real user navigates the app.
- The Septim sub-agent structure maps directly to Joe's Batch/Chef orchestration: named roles, shared project context, no re-briefing between calls.
- Prompt Master's approach — detect tool, extract intent across 9 dimensions, max 3 clarifying questions — is a clean framework for structuring JoeFlow's rewriter prompts.
Terms worth knowing.
- Claude skill
- A reusable instruction package you install into Claude that gives the assistant a specific capability or workflow, activated automatically or by name during a chat.
- skill.md file
- The markdown file at the core of a Claude skill that defines its instructions, triggers, and behavior, written in plain text rather than code.
- Skill Creator
- A pre-installed Claude skill that interviews you about a repeated workflow and generates a complete reusable skill file, then runs side-by-side tests with and without the skill active.
- Prompt Master
- A community-built Claude skill that rewrites messy brain-dump prompts into structured, optimized prompts tailored to whichever AI tool will receive them.
- Fact Checker skill
- A Claude skill that scans any text for factual claims and cross-references each one against external sources via web search, returning a report of confirmed, unverifiable, and false statements.
- Humanizer skill
- A Claude skill that detects and removes telltale patterns of AI-generated writing, optionally matching the rewrite to a sample of your own writing style.
- Em dash overuse
- A common tell of AI-generated text where the em dash punctuation mark appears far more often than a human writer would naturally use it.
- Rule of three
- A writing pattern where ideas or descriptors are grouped in sets of three, frequently overused by AI models to the point of becoming a recognizable signature.
- Claude Code
- Anthropic's command-line coding tool that lets Claude operate directly inside your project, reading and editing files, running commands, and executing multi-step engineering tasks.
- Playwright
- A browser automation framework, packaged here as an official Anthropic skill, that lets Claude drive a real web browser to click through flows, fill forms, and capture screenshots for end-to-end testing.
- Superpowers skill
- A community Claude Code skill that enforces a disciplined build loop — planning before coding, writing tests alongside code, and self-review — instead of letting the agent ship unchecked output.
- Plugin marketplace
- The in-app catalog inside Claude Code where skills, agents, and other plugins can be browsed and installed without manually downloading files.
- Subagent
- A specialized helper agent spawned by Claude Code with its own role, instructions, and isolated context, allowing different perspectives to operate on the same project without polluting the main session.
- Septum Agents Pack
- A bundle of ten named, role-specific subagents — covering planning, architecture, branding, marketing, finance, design, legal, customer experience, research, and coordination — that operate as a virtual team inside one project.
- /review
- A Claude Code slash command that reviews recent code changes for bugs, edge cases, security issues, and design problems before they get merged.
- /security-review
- A Claude Code slash command that audits pending changes on the current branch specifically for security vulnerabilities.
- Context window
- The fixed amount of text a language model can hold in active memory during a session, including instructions, prior messages, and tool outputs.
- /context
- A Claude Code slash command that shows what is currently consuming the session's context window so you can see where memory is being spent.
- /compact
- A Claude Code slash command that compresses the current conversation into a shorter summary, freeing space in the context window while preserving the parts you flag as important.
- Auto-compact
- Claude Code's automatic behavior of summarizing a session when the context window fills up, which can drop important details if it happens without user direction.
Things they pointed at.
Lines you could clip.
00:42
“Most are broken, outdated, or just system prompts with a fancy name.”
02:46
“That step is the difference between a skill that works and a skill that you think works.”
08:40
“By default, you are the QA department.”
15:00
“That's not a prompt. That's institutional memory.”
15:21
“A skill is just a tool. A tool you never improve is just clutter.”
15:57
“The difference between a calculator and a full-on computer.”
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
00:00So most people use Clawd in the most basic way that you can imagine. They open up a chat, they type in a prompt, they get an answer, and then what do they do?
00:10They repeat it. And that's basically it. They never actually go deeper, but Claude is actually capable of so much more than that, and most people have no idea.
00:21With Claude skills, you can catch bugs before you even see them. You can keep your sessions clean for hours and you can make Claude behave exactly the way that you needed to for a specific situation. And the best part is it gets smarter every time you use it.
00:37There are now thousands of Claude skills and I have tested more than 250 of them. Most are broken, outdated, or just system plumps with a fancy name.
00:48But buried inside that mess are nine skills that represent a completely different way of using Claude. So I'm splitting them into three categories for today's video. Essential Claude skills that you can use right now, Claude code skills for building, and also one co work pattern that ties everything together.
01:09This is gonna be a fun one. Don't forget to check out the three score community down in the description below. But right now, let's get into it.
01:16So we are starting off with category number one and that is the essential Claude skills. So these work in Claude desktop. There's no terminal, no code needed.
01:25Just upload a zip file through customize and you are set. So let's get into skill number one and that is skill creator. This is our first skill and it's basically something that sounds pretty funny, but it actually solves most of your problems.
01:40This is one that is built for beginners. But ironically, it's one of the most overlooked skills that you will find anywhere.
01:47Not because it's hard to find, I mean, it's actually pre installed, but most people just never think to actually use it first. So you don't need to download anything, just open up Claude desktop and then go to any tab in Claude and then just type in this prompt. I want to create a skill using slash skill dash creator.
02:05Just type that in. And here's the problem that most people actually run into. They have a workflow that they just repeat about three, four, five times a week.
02:14I mean, it's the same instructions, the same format, the same corrections, and every time they start a new chat, they explain it all over again from scratch. Skill Creator actually lets you turn that into a reusable skill in just a few minutes.
02:30You describe what you want the skill to do, and then Claude actually interviews you with a few questions just to basically, like, nail down all of the details, and then it builds the full skill dot m d file for you, and there's no markdown editing. It's just a pure conversation. But here's the part that nobody actually shows.
02:49After it builds the skill, it automatically generates test cases with the skill active and also without the skill. So you can actually see the output side by side. So if the skill doesn't actually improve the result, then you know before you ever even commit to using it.
03:04That step is . I mean, most people skip it, so I'm telling you don't. Now we get on to the second skill and that is prompt master.
03:17So this one is on GitHub. All you have to do is basically just go and search for prompt master by n dash I dash d dash h dash I dash n dash j dash s.
03:30And then what you do is you download the repo as a zip file, and then you go over to Claude, you open up the sidebar, and then you click customize and then skills, and then upload a skill, and you are done. So here's the problem. Now most people plant Claude with messy brain dumps.
03:48It's always long. It's unstructured, and it's basically just full of vague instructions.
03:53And for simple tasks, that is fine, but for anything that's complex, the the output quality drops pretty fast and it's noticeable.
04:02So prompt master actually takes whatever you throw at it and it turns it into a structured optimized prompt according to the best practices. It detects which AI tool the prompt is for, it extracts your intent across nine dimensions, and it asks a maximum of three clarifying questions if something critical is missing from all of this, and it delivers one clean prompt that you can simply just copy and you can use.
04:29It works with over 30 tools Claude, ChatGPT, Cursor, Midjourney, Stable Diffusion, you name it, all of it is there. But the real power is using it inside Claude itself.
04:40You write your messy brain dump, you add, use the prompt master skill on this one and then Claude actually rewrites it before processing. One extra step but dramatically better outputs.
04:53Now here's a quick pro tip, you can actually add this to your general instructions inside Claude desktop settings so it auto enhances on every complex prompt. You never have to think about it ever again. Now that takes us perfectly into skill number three and that is the fact checker.
05:08So multiple versions of this exist. The easiest one is to grab from Daymaid's Claw Code Skills repo on GitHub and all you do is just download, zip, upload through, customize, and then it's basically the same process that we discussed before. The AI generates confident sounding text.
05:23So sometimes it is wrong and if you're putting a name on it like a LinkedIn post, a script, or a client deliverable, then wrong is basically a huge problem. So this skill runs a systematic fact verification pass on every or any text that you give it.
05:41So it reads every factual claim and it cross references against external sources using web search and it gives you a report like what's confirmed, what's unverifiable, and what what's flat out false.
05:54So I use some three things, anything AI generated that goes out publicly, so my own video outlines before I film and other people's content when I want to check if what they are claiming holds up. Now here's another pro tip for you, Once you have it installed, just add a step to your writing skills that runs the fact checker automatically before giving you the final output.
06:15You stop the problem before it even reaches you. Now we go on to skill number four and that is the humanizer. Grab this from Blader's humanizer repo on GitHub, download it as a zip and upload through customize and takes about thirty seconds, and this one is pretty straightforward.
06:32It removes the signs of AI generated writing from any text that you give it, and you should be running this on anything AI generated that other people are going to read over. It's based on the Wikipedia's comprehensive signs of AI writing guide and it detects 25 specific patterns, inflated symbolism, promotional language, m dash overuse, and I know you know what I'm talking about, and also the rule of three AI vocabulary words like leverage and streamline, all those vague attributions and the filler phrases, all of those things that make the AI text feel like an AI text.
07:13But here's what makes this one better than just telling Claude to make it sound more human. If you give it a sample of your own writing, then it actually analyzes your sentence rhythm, your word choices, your quirks, and then applies those to the rewrite instead of producing generic clean output.
07:31I run this on LinkedIn posts, client emails, onboarding docs, community guides, anything that goes out with my name on it. And if you use writing skills like LinkedIn writer or a newsletter skill, add a humanizer as a final step so that it runs automatically before you see the output.
07:51That's only gonna give you the best results. Now we are moving on to category number two and that is the Claude code skills. So those four work in any Claude chat.
08:01But if you use Claude code, then the next four skills operate at a completely different level. For skill number five, I have a powerful combo of playwright and superpowers. This combines two skills.
08:14Playwright is an official Anthropic skill, so you can find it in the Anthropic skills repo on GitHub, and Superpowers is from Obera's GitHub repo.
08:25So you can install both through the plug in marketplace in the code tab, and you'll need Playwright installed on your machine. Super important.
08:34Everyone shows Claude writing code, but nobody actually shows Claude testing the code that it actually just allowed. Claude bolds you a web app, and it looks right, but did it test the sign up flow?
08:48Does the button work on mobile? You don't know. By default, you are the q and a department.
08:55Playwright handles the testing side. Once active, Claude navigates your local app like a real user. It clicks through flows, it fills out forms, it captures screenshots at every step, and it reports back exactly what broke.
09:08But superpowers takes it even further. It forces Claude to plan before it codes, write tests before it ships, and it even reviews its own work before you ever see it.
09:20So instead of Claude building something and testing it after, it's building with quality gates baked into every step.
09:30Now the moment that actually sold me on this was that Claude built a simple app. It planned the architecture first, it wrote tests alongside the code, and then it then played right to verify everything end to end. And it found a button that was technically there but visually hidden on screens under 400 pixels wide.
09:49That's the kind of bug that reaches your users if nobody catches it first. Now we have to go to skill number six and that is Septum Agents Pack.
10:00So this is going to take you to the next level. Most people using Claude Co do everything with one agent in one context window. By default, you are the QA department.
10:11That's like hiding one person to be your architect, your lawyer, your CFO, your marketer, and your QA tester all at once.
10:20This skill drops a full name team in your project in just a single move. 10 specialized sub agents each with a name and a role.
10:31Atlas handles the planning, Luca handles the architecture, Canon handles the brand, Amber handles the marketing, and Telly handles the finance, Nova handles the design, Ward handles the legal side, and Meta handles the customer experience. And then on top of that Juno handles the research, PIP handles the coordination, so you genuinely have an entire team.
10:53In practice, you are building a new feature. You ask Claude to use Atlas to plan it out and then Ward to check for the legal issues and then Tally to estimate the costs. That's three perspectives.
11:07That's three areas of expertise in just one project. You never have to re explain context before each agent reads from the same project files.
11:17Now the original Septum Agents pack costs about $49, but I built my own free version that does the exact same thing, and you can actually grab it in the free school community. The link is down in the description below.
11:30Now for skill number seven, and that is the view, and this has to be one of the most underrated skills. Claude writes code, but it doesn't really view its own work unless you tell it to. You type in slash review after any bold and then Clawd actually checks for bugs, edges, cases, security issues, and design problems all for you.
11:51The stuff a careful human reviewer would catch and you can actually use it as your last step before merging anything. So it runs locally, it costs nothing extra, and it catches things that Claude missed the first time around.
12:04But did you know that you can actually do a security review directly in code? All you have to do is just type in slash security dash review and you can see a review of the pending changes on your current branch.
12:17Now let's discuss skill number eight and that is context and compact. Now every tool under Claude dumps raw data into your context window. So after about thirty minutes of real work, half your window is noise and when it fills up, then Claude auto compacts and it forgets what it was doing.
12:36So slash context shows you what's eating your window up right now and slash compact lets you compress it on your own terms instead of letting Claude decide randomly. So the habit that you're going for here is you wanna run context every twenty minutes and then when you hit 60, you run compact with a focus like compact focus on auth module and then your sessions run longer and then Claude stops forgetting and you stop wasting prompts at explaining things that you really don't need to.
13:05So that was category number two and that was basically just the code skills. Now we are getting into category number three, and that is the co work pattern. Okay.
13:14And let's start off with skill number nine, and that is self improving skills. So we are wrapping up our list, and we have our final learning dot m d pattern.
13:26So this is the one that I tell you to implement today because it's so important and that's basically if you only if you only had to take away one thing from this video. Now there's no install needed. This is just a pattern that you apply to any skill that you already have and you just create a file called learnings dot m d alongside your existing skill file and then you add in one line to your skill that says before running the learnings dot m d.
13:53And here's why. You build a skill and it works okay, but it makes the same mistakes every time. You correct it and then the next run and it's just the same mistakes.
14:02The learnings dot MD pattern fixes this for you. So after every run, Claude logs what worked, what didn't, and what you corrected. They did specific scope to that skill.
14:12So the next time the skill runs, it divides those notes first and the skill actually compounds. So I run a writing skill on day one, and then the output is decent but generic. You corrected and then the intro was too long, the headlines needs a number, the client hates passive voice, those corrections go into learnings.md.
14:35And by then five, the skill has internalized all of that.
14:40The learnings file from one of my content skills says something like headlines with numbers consistently outperformed, intro paragraphs over 80 words flagged as too long by client, avoid the word leverage.
14:55Client has specifically rejected this three times. And that's not a prompt. That's institutional memory.
15:03That single change turns any static skill into a system that learns your preferences over time. And here's the thing that nobody says out loud about Claude skills.
15:16Installing a skill does not make you better at using Claude. A skill is just a tool. A tool that you never improve is just clutter.
15:27The one pattern that separates every skill on this list from everything else that you've installed is the learnings file. Not because it's the most impressive skill, but because it's the only one that gets better the more you use it.
15:44Every correction that you make, every preference that you log, every mistake that Claude repeats and you catch, that all goes in and it compounds over time. Now that's the difference between a calculator and a full on computer.
15:59So that is all nine skills and every install link is all down in the description below in the free school community. Go and check it out and I will catch you on the next one.
The Hook
The bait, then the rug-pull.
Most people open Claude, type a prompt, get an answer, and repeat. Zinho Automates tested over 250 published skills so you don't have to — and the verdict is blunt: most are broken, outdated, or just system prompts with a fancy name. What follows are the 9 that actually change how Claude behaves.
Named ideas worth stealing.
00:46list
Three Categories of Claude Skills
- Essential (Claude Desktop, no code)
- Claude Code Skills
- Co-Work Pattern
The organizing framework for the whole video — separates no-code accessible skills from CLI-power skills, with a meta-pattern that applies to both.
Steal forAny listicle or course structure where you want to segment by skill level without alienating either end of the audience
13:16concept
learnings.md pattern
Add a learnings.md file alongside any skill. One line in the skill reads it first. Claude logs corrections after every run. Skills accumulate institutional memory instead of resetting.
Steal forAny repeating Claude workflow Joe has — dictation corrections for JoeFlow, content formatting prefs for MCN, output standards for Batch templates
02:46concept
Test-Before-You-Commit (Skill Creator's side-by-side eval)
After building a skill, Skill Creator auto-runs it and the non-skilled equivalent side-by-side. If the skill doesn't improve the result, you know before committing.
Steal forAny A/B prompt evaluation workflow
How they asked for the click.
VERBAL ASK
15:59link
“Every install link is down in the description below in the free Skool community.”
Soft close — no pressure, purely additive. Community CTA was also seeded at 01:08. Clean execution, not pushy.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
