Every new AI session with Claude, you start from zero.
You re-explain your audience. You re-paste your style guide. You retype the same 400-word context block you’ve been copying from Notion since January. And when you forget one piece of it, which you always do, the output misses the mark entirely.
TLDR; Claude Skills are reusable SOPs for your AI work. Write the instruction once, install it once, and Claude can execute them on demand without re-explanation. As of April 2026, Skills 2.0 ships with a built-in eval loop (a quick automated test that runs your skill on sample inputs and compares outputs) before you approve the skill. The result: consistent, contextual outputs even after weeks away from a workflow.
Ilia Karelin, creator of the Prosper in AI newsletter, had the same start-from-zero problem. Before every research session, he’d spend time re-explaining the same frameworks, the same audience context, the same lens through which he analyzes AI trends. Not because the work was hard, but because Claude couldn’t remember it.
In a recent session on One Shot Show Episode 7, Ilia joined me and Wyndo to walk through a better approach: Claude Skills 2.0. Not the watered-down version most people use, but the full eval-driven system documented in Anthropic’s official Skills guide that lets you install a workflow once and invoke it forever, including from your phone.
By the end of this article, you’ll understand what Claude Skills are (most people confuse them with prompts), how to decide when to build one, the three levels of skill complexity with real examples, and how Ilia fires a production-grade research skill from his phone while his laptop saves the output automatically.
👋 Julley, I'm Dheeraj, an AI systems builder.
I build production-grade AI systems at work by day and ship my own products by night, 9 and counting, including SubflowAI and the Content OS Agents Toolkit. This newsletter is the bridge between those two worlds. Every system, every build, documented step by step.
Join 1,800+ builders getting the exact AI setups, prompts, and workflows that actually work in your business.
Why Does Restarting From Zero Cost You So Much?
The hidden tax on solo creators isn’t the hard work. It’s the repetitive setup before every AI interaction.
Wyndo described his pre-skills workflow on the show: elaborate prompt templates saved in Notion, copied into each new session, updated when they drifted out of date. After adopting skills, that whole template block became unnecessary. Triggering a content repurposing run, in his words, takes a single phrase like “repurpose this” instead of a copied prompt block.
I tracked this more precisely after running his own skills on n8n workflow creation.
Before Skills 2.0, his outputs were what he called “okay-ish.”
After running the 2.0 eval on his n8n skills, quality jumped to roughly 95%.
I now rarely logs into the n8n interface directly. The Claude Code n8n MCP and n8n Claude skills handles my n8n workflows generations.
The math on this adds up fast. If you spend 20 minutes re-explaining context before each major AI task, and you do that task four times per week, that’s 80 minutes of pure setup overhead every week. Skills eliminate it.
What Is a Claude Skill, and How Is It Different From a Prompt?
A Claude Skill is a reusable instruction file, written once and installed into Claude’s skill library, that fires on demand without any re-explanation.
Skills are AI-model-agnostic: Claude, OpenAI, Gemini, and Grok all support the same concept, though the implementation differs slightly.
The key distinction Ilia draws at the start of the session (4:45) is the diagnostic question: “What am I doing too often manually? A couple of times a week, it takes me a couple of hours every time I do it. That’s usually kind of the good baseline.”
That question filters skills from prompts. A prompt is a one-time instruction. A skill is an SOP you’ve decided is worth codifying.
Skills vs. Agents vs. Commands: The Distinction That Matters
Most people conflate these three, and the confusion leads to building the wrong thing. I draw the clearest distinction at 8:43, using an employee analogy:
“Agents are like employees with their own domain of knowledge. Skills are the domain expertise you arm them with.” - Dheeraj
In practical terms: an agent has its own context window, operates independently on a specific task, and then returns results to your main conversation. A skill is a module that runs within your current session. You can arm an agent with skills, or use skills directly in any conversation.
Anthropic deprecated slash commands, the older /commands directory approach. Skills 2.0 replaces them. The commands directory still exists for legacy support, but new workflows should use Skills.
One technical detail worth knowing: Skills 2.0 introduced two flags, user_invocable (you trigger it manually) and system_invocable (Claude can trigger it on its own when it detects a match).
Setting user_invocable: true and system_invocable: false means the skill fires only when you explicitly invoke it. As I explain at 11:31, these flags matter when you have skills with overlapping trigger conditions and don’t want Claude auto-firing the wrong one mid-session.
Commercial break: Claude Code Builder cohort
The founding batch of my Claude Code cohort starts in July on Maven. Four weeks, six sessions. Learn to build the AI system with me.
Only 12 Seats. When they’re gone, the founding price ($797) closes and Cohort 2 opens at $1,597.
Use code GENAI20 for 20% off. Expires July 9. Check the Syllabus →
What Are the Three Levels of Claude Skill Complexity?
Skills fall into three levels of depth, and choosing the right level for each workflow is one of the most important decisions you’ll make.
Level 1: skill.md with name and description only
Best for: Simple, standalone instructions that don’t need context
Contains: Name, description, basic trigger phrase
Example: A formatting skill that applies consistent heading styles
Level 2: Full instructions + output format + constraints
Best for: Multi-step workflows with specific output requirements
Contains: Detailed instructions, output format specification, edge case handling
Example: A social media repurposing skill that converts newsletter content into LinkedIn posts with specific formatting rules
Level 3: skill.md + reference docs + scripts + templates
Best for: Context-heavy workflows tied to your specific audience and brand
Contains: skill.md + reference documents about your philosophy/audience + test output examples + optional Python scripts
Example: Ilia’s fast research skill (covered in detail below)
One important constraint that Wyndo and I both noted: skills that include code files, such as Python scripts or bash commands, can only run in Claude Code (the terminal environment) and Claude Cowork. They cannot run in Claude (the chat/projects interface). Pure-text skills work in both.
Both Wyndo and I independently mentioned having 30+ skills in their folders. That accumulation happens fast once you internalize the pattern.
I am also clear about the anti-pattern: “If you have to repeat something like a prompt, a long prompt three times a week, that’s a pattern. Then I would go and create a skill out of it.” Below that threshold, you’re creating overhead.
If you run four major AI tasks a week at 20 minutes of context setup each, that is 80 minutes of pure overhead every week. Not output. Retyped context.
PluggedIn has the skill templates and eval scaffolding from this article so you can install a working skill in 15 minutes instead of building from scratch.
How Do You Build a Claude Skill Using Skill-Creator and AskUserQuestion?
The fastest path to a working skill is the skill-creator skill, which ships with Claude Code and generates other skills using a built-in eval loop. Here’s the process Ilia walked through live in the session, including building a “second opinion” skill from scratch.
Step 1: Identify a skill candidate with the 3x/week rule.
Ask yourself what you’re doing too often manually. If the same prompt pattern appears more than three times a week, it’s a skill candidate.
Step 2: Decide on your complexity level.
Simple one-off instructions = Level
Workflows with output format requirements = Level
Workflows tied to your specific content, audience, or philosophy = Level 3.
Step 3: Launch skill-creator explicitly.
In Claude Cowork, type /skill-creator or say “I want to create a skill using skill-creator.” Both Ilia and I emphasized using the explicit invocation, not asking Claude to “create a skill” without naming skill-creator. Without the explicit invocation, Claude may guess at the framework rather than loading the skill-creator infrastructure.
Step 4: Let AskUserQuestion interview you.
The AskUserQuestion form, which appears as an interactive form at the bottom of the Cowork prompt input, will ask up to three multiple-choice questions. For the second opinion skill, it asked about the goal, input types, and desired perspectives. Claude interviews you rather than making assumptions.
Step 5: Review the eval results before approving.
This is the step most people skip. Skill-creator generates roughly six test cases: three with the skill active and three without. Reviewing the difference tells you whether the skill is doing what you intended or only adding overhead.
Wyndo’s framing at 22:01 sets the right expectation:
“Your time is spent not building the skill. Your time will be spent around evaluating the skill output.” - Wyndo
The second opinion skill Ilia built live critiques writing from four perspectives: support the argument, destroy the argument, identify logical flaws, and propose alternative framings.
The inspiration came from Andrej Karpathy’s approach: asking an AI to both support and destroy an opinion simultaneously. The results are often uncomfortably accurate.
Step 6: Define your trigger phrase.
Natural language trigger phrases are set during the build. For the second opinion skill, phrases like “second opinion on” and “poke holes into” would fire it. This is more flexible than slash commands and works with voice-to-text tools like Wisprflow.
What Does a Production-Level Claude Skill Look Like?
Ilia’s fast research skill, which he walks through at 43:53, is the clearest example in the session of what a mature, production-grade skill looks like.
It’s a Level 3 skill. The folder contains:
skill.mdwith name, description, and trigger phrasesA commands file with specific research steps
Reference documents containing Prosper newsletter’s philosophy, audience beliefs, and the specific lens through which Ilia frames AI concepts
Test output examples showing Claude what “good” looks like for this specific newsletter
The skill runs a web search, extracts relevant information, frames it through the Prosper audience lens with a “possible angles” section, and saves the output as a .md file directly to a local newsletter folder on Ilia’s laptop. The file is immediately ready for his draft pipeline.
The reference documents are what separate Level 3 from Level 2. They contain entries like: what the Prosper audience believes about AI, what kinds of arguments resonate with them, what Ilia’s editorial philosophy is.
These aren’t general instructions. They’re the context that makes outputs contextual rather than generic.
“Reference docs can contain your newsletter beliefs, audience context, and personal philosophy. this makes outputs contextual, not generic.” - Ilia Karelin (at 43:53)
When I asked whether Ilia uses Perplexity or Tavily MCP servers to enhance the research, Ilia confirmed he relies on Claude’s built-in web search. The reference docs do more to improve output quality than adding external search tools.
How Do You Fire a Claude Skill From Your Phone - Claude Dispatch?
This is where the session’s most memorable moment happens. At 46:38, Ilia picks up his phone and types: “fast research GPT 5.5 versus Opus 4.7.” He clicks send. Instantly, the research appears on his laptop.
The skill fires, runs the web search, formats the output, and saves a .md file to his newsletter folder, all from a phone tap.
The feature making this possible is Dispatch, which lives inside Claude Cowork (not a separate product).
How Claude Dispatch works:
Open Claude Cowork on your laptop
Find the Dispatch option in the interface
Scan the QR code with your phone
Send messages from your phone to your active Cowork session
Results appear on your laptop in real time. Outputs can save files directly to your local folders. Your laptop must remain awake for the connection to persist.
At 47:31 I compared Dispatch to OpenClaw, another tool for remote Claude access covered in Episode 5 of the show. The difference: OpenClaw requires multi-hour configuration. Dispatch requires one QR code scan.
The implication for solopreneurs: your phone becomes a remote control for your entire AI workspace. You can trigger research, content repurposing, or any installed skill while you’re away from your desk, and the output waits for you when you return.
Dispatch vs. OpenClaw comparison:
Dispatch (Claude Cowork)
Setup: One QR code scan
Works with: Claude Cowork only
File output: Yes, saves to local laptop folders
Laptop must stay awake: Yes
OpenClaw
Setup: Multi-hour configuration
Works with: Claude Code terminal
File output: Yes
Laptop must stay awake: Yes
What Changed in Claude Skills 2.0?
Skills 1.0 was a skill.md file. It worked, but outputs were inconsistent and there was no systematic way to test whether the skill was improving anything.
Skills 2.0, as Wyndo explains at 33:44, introduced four significant changes:
The skill-creator skill with built-in eval and test case generation
The AskUserQuestion interview form that reduces prompt ambiguity
Automatic test cases (approximately six per skill: three with, three without)
Auto-save, ending the previous workflow of downloading a zip file and re-uploading it
The eval loop is the biggest change. Wyndo re-ran the eval on all 30+ of his existing skills after the update shipped. I described my n8n workflow creation skills going from “okay-ish” to approximately 95% quality after running the 2.0 eval.
The update was also retroactive. You can improve old skills built under 1.0 by running them through the 2.0 eval, without rebuilding from scratch.
One structural detail about skill availability: skills created at the user level (home directory) are available across all projects and your entire system. Skills created inside a specific project are scoped to that project only. When building a skill you’ll want everywhere, tell Claude to “create this skill at the user level.”
Claude Desktop Chat and Claude Code (terminal) are also separate skill ecosystems. Pure-text skills added via Customize → Skills appear in chat. Skills with code files live in Claude Code only.
Get PluggedIn
Stop retyping the same 400-word context block before every Claude session.
Every week without skills installed is another 80 minutes of setup overhead you already did last week.
Get PluggedIn to go from copying context into every new Claude session to triggering a production-grade skill with a single phrase
Frequently Asked Questions About Claude Skills 2.0
Are Claude Skills available across all my projects?
It depends on where you create the skill. A skill built at the user level, in your home directory, is available across all projects and your entire system. A skill built inside a specific project is scoped to that project only. When you want a skill everywhere, tell Claude to create it at the user level explicitly.
Can I use Claude Skills in chat or only in Claude Code?
Pure-text skills work in both chat and Cowork. Add them via Customize → Skills in Claude Desktop. Skills that include code files, Python scripts, or bash commands require Claude Code in the terminal and cannot run in the chat window. Claude Desktop and Claude Code are separate skill ecosystems.
What is the difference between a Claude Skill and an agent?
Agents have their own context window, operate independently on a specific task, and return results to your main conversation. Skills are domain expertise modules that run within your current session. You can arm an agent with skills, or use skills directly without any agent involved.
When should I build a skill instead of using a prompt?
Use the 3x/week rule. If you’re repeating the same prompt pattern more than three times per week, convert it to a skill. Below that threshold, the overhead of maintaining the skill outweighs the benefit. Too many rarely-used skills create cognitive overhead about what each one does.
What changed in Claude Skills 2.0?
Skills 2.0 introduced: built-in eval with automatic test case generation (three with skill, three without), the AskUserQuestion interview form to reduce prompt ambiguity, and auto-save. The eval loop is the biggest improvement. It’s retroactive, so you can run it on skills built under 1.0 to upgrade them.
How do I fire a Claude Skill from my phone?
Use Dispatch inside Claude Cowork. Pair your phone with a single QR code scan, then send text messages from your phone to trigger skills in your active Cowork session. Results appear on your laptop in real time. Files can save directly to your local folders. Your laptop must remain awake.
How many Claude Skills should I have?
Only build skills for workflows you repeat three or more times per week. Power users like me and Wyndo have 30+ each, but each skill in the folder should earn its place. Quality and specificity matter more than volume. Too many underused skills create confusion about what fires when.
Key Takeaways
The 3x/week rule filters real skills from noise. If you repeat the same prompt more than three times per week, it’s a skill candidate. Below that, it’s overhead.
Building time is short. Eval time is long. Most of your effort on a new skill should go into reviewing the test cases, not writing the instruction. This is where quality is won or lost.
Level 3 skills with reference docs produce contextual outputs. Including your newsletter philosophy, audience beliefs, and test output examples inside the skill folder makes outputs specific to your work, not generic.
Skills 2.0 is retroactive. If you have existing skills built under the 1.0 system, run them through the 2.0 eval to upgrade them without rebuilding.
Dispatch turns your phone into a skill remote. One QR code pairs your phone to your active Cowork session. You can trigger any installed skill from anywhere your phone has signal.
User-level vs. project-level scope matters. Skills built at the user level are available everywhere. Skills built inside a project are scoped to that project. Build globally when you want a workflow available across all contexts.
Claude Code and Claude Desktop are separate ecosystems. Pure-text skills work in both. Skills with code files live in Claude Code only.
Your 15-Minute Challenge
Pick one prompt you’ve used more than three times this week. Open Claude Cowork. Type /skill-creator and follow the AskUserQuestion interview. When the eval runs, compare the test outputs: three with your skill, three without. If the skill outputs are noticeably more consistent and contextual, approve and save it at the user level.
That’s the full loop. One skill. Fifteen minutes. Your success criteria: the next time you need that workflow, you trigger it with a single phrase instead of retyping the context block you’ve been copying from Notion.
For the full walkthrough including Ilia’s live build and the Dispatch phone demo, watch One Shot Show Episode 7. The skill-creator demo starts at 23:09. The phone firing moment is at 46:38. Both are worth watching at least once to see the full UI flow.
Connect With the Hosts and Guest
Ilia Karelin: Prosper in AI on Substack
Dheeraj Sharma: GenAI Unplugged on Substack | YouTube
Wyndo: The AI Maker on Substack
Your PluggedIn assets for this post
What’s inside:
05-example-level3-research-skill.pdf - Example Level3 Research Skill
06-example-skill-eval-comparison.pdf - Example Skill Eval Comparison
01-skill-candidate-audit.md - Find Your Best Skill Candidates
02-skill-creator-prep.md - Skill-Creator Interview Prep
03-skill-eval-reviewer.md - Read Your Skill Eval Results
04-skill-design-worksheet.md - Claude Skill Design Worksheet
Coming soon on One Shot Show: a deep dive on memory management in long Claude Code sessions, including Dream Mode and the 1M context window.
