Claude Skills Built Me an AI Agent Army (They Run Everything Now)

6mo ago

In this episode, Amir takes us through how to use Claude Skills to build digital employees. We cover practical demos including an A/B testing idea agent, marketing insight analyzer, and a live build of a tweet-to-newsletter converter. You'll learn what Claude Skills actually are, why they represent the biggest leap since sub-agents, and how to build them yourself, even if you've never written a custom AI workflow before. Timestamps 00:00 – Intro 01:05 – What are Claude Projects 02:40 – Sub-agents in Claude Code explained 03:34 – Introducing Claude Skills 05:58 – Context rot and the performance degradation problem 08:01 – Why Claude Skills is Great 11:08 – Building a UTM link generator with Artifact Builder 17:41 – Claude Skill Demo: A/B test generator for website optimization 20:32 – Claude Skill Demo: Marketing analytics insights from campaign data 23:40 – Building a Claude Skill: Creating a tweet-to-newsletter converter skill 30:32 – Final Thoughts on Claude Skills 30:58 – Why AI adoption is falling and how better prompting solves it Key Points * Claude Skills are automated workflows that apply globally or per-project, pulling context only when relevant to specific tasks * Skills solve the "context rot" problem where too much context degrades LLM performance and increases hallucination * You can create custom Skills using markdown files with instructions, reference documents, and executable scripts * The tweet-to-newsletter converter built live demonstrates Skills' ability to match tone and style with minimal training * Poor AI fluency and prompting (not the tools themselves) explain why enterprise AI adoption is declining Section Summaries 1. Claude Projects vs. Sub-Agents vs. Skills Projects are collaborative workspaces with custom instructions, context files, and memories shared across teams. Sub-agents (in Claude Code) handle complex multi-step tasks by delegating to specialized agents within a single conversation. Skills sit above both as reusable, automated workflows that load context selectively and can execute custom scripts for deterministic results. 2. The Context Rot Problem Research shows that adding excessive context to LLMs can degrade performance and trigger hallucinations, not improve accuracy. Skills address this by pulling reference files only when the task requires them, avoiding the "bombarding your coworker with every document" problem that plagues traditional prompt engineering. 3. How Skills Work Under the Hood Skills use a markdown file defining the task, instructions, and workflow steps. They can reference additional documents (brand guidelines, glossaries) and run custom Python scripts for deterministic data analysis. This three-layer approach 1) instructions, 2) references, 3) code. This gives users precise control over outputs without relying on the LLM's judgment alone. 4. A/B Test Generator for Conversion Optimization Amir demonstrates a custom skill that scrapes a website URL and generates prioritized A/B test ideas using the ICE framework (Impact, Confidence, Ease). The skill recommended moving social proof above feature descriptions, a suggestion Amir implemented immediately because it was grounded in conversion best practices. 5. Marketing Analytics Without Hallucination The traffic analytics demo shows Skills running Python scripts on campaign data to produce accurate performance breakdowns by channel, spend, and conversion. Unlike Projects where the LLM interprets data non-deterministically, Skills use predefined calculation scripts and reference a metrics glossary, drastically reducing hallucination risk. 6. Live Build: Tweet to Newsletter Converter Greg requests a skill that turns his viral tweets into newsletter drafts matching his writing style. Using the Skill Creator skill (meta!), Amir uploads example tweets and newsletters as reference files, generates the skill package, and produces a polished newsletter draft on the first try, complete with tone matching and structural