NemoClaw Knowledge Wiki

Using Claude Code to analyse LinkedIn data

18 min read

YouTube Summary Report

Generated: 2026-03-07 · API: Gemini 2.5 Flash · Modes: Summary, Key Timestamps

Video 1

URL: https://youtu.be/D4ZL6Ld51\_s?si=g7O1jNCuazs6ytLj

Summary

This video showcases a powerful application of AI to LinkedIn data, turning a seemingly flat professional network into a dynamic tool for career advancement and job searching. The main topic revolves around exporting your comprehensive LinkedIn data, processing it with an AI tool like Claude Code, and then leveraging the resulting insights for strategic career development, particularly in identifying job fit, skill gaps, and crucial networking opportunities. The speaker, a cybersecurity professional, begins by highlighting a seemingly insignificant connection with “Cheryl Auger” on LinkedIn, only to later discover through AI analysis that she holds a key to addressing a potential career gap. He then recounts his own LinkedIn journey, illustrating how his network saw minimal growth during his “W2 stuff” phase but exploded after he began building a personal brand, consistently delivering value, and engaging publicly through his “Simply Cyber” initiative. This personal experiment, backed by visualizations of his cumulative connections and monthly growth rates, emphatically validates his long-held advice: consistent content creation and networking directly correlate with significant professional network expansion. The “game-changer” aspect of the AI analysis emerges when the speaker feeds his LinkedIn data alongside a specific job description (a Director of IT role at a water utility) into Claude Code. The AI performs a detailed gap analysis, not only highlighting his strengths (e.g., PhD in Cyber Ops, GRC experience, strong communication) but critically identifying precise areas where his experience is lacking for that particular role (e.g., SCADA/ICS experience, utility sector knowledge, general IT ops vs. security-focused). Beyond mere identification, the AI then cross-references his entire LinkedIn network and provides a prioritized list of connections who possess the expertise to help close these gaps, with Cheryl Auger topping the list for SCADA/ICS knowledge. The conclusion emphasizes the profound utility of this AI-driven approach. It allows professionals to transform their often-underestimated LinkedIn network into a “gold mine,” providing actionable insights that would be near-impossible to derive manually. The tool offers not only a clear understanding of personal strengths and weaknesses relative to desired roles but also a strategic roadmap, including who to contact for informational interviews, how to adjust a resume to emphasize relevant experience, and key talking points to prepare for interviews. This method provides an unprecedented level of personalized career guidance, making hidden connections visible and empowering individuals to proactively shape their career paths with targeted networking and skill development.

Key Timestamps

Here are the key moments and topics in the video with timestamps:

  • 00:00 — Introduction: The speaker introduces Cheryl Auger, a LinkedIn connection, and hints at the hidden potential within LinkedIn data for career advancement.
  • 00:06 — LinkedIn Data as a “Gold Mine”: The speaker presents the core idea that personal LinkedIn data can reveal valuable career opportunities and connections.
  • 00:27 — Discovering a Data Visualization Method: He explains how he found Logan Currie’s blog post, which detailed how to export LinkedIn data and analyze it with Claude Code for visualizations.
  • 00:35 — Speaker’s Career Advice and Personal Branding: He shares his background in cybersecurity and his consistent advice about building a personal brand, delivering value, and growing a network.
  • 00:48 — Experimenting with His Own LinkedIn Data: Decides to run the same experiment as Logan Currie to validate his personal branding advice with data.
  • 01:02 — Network Growth Over Time Visualization: Displays a graph showing his LinkedIn connections, highlighting a significant “hockey stick” growth after launching “Simply Cyber” in 2019 and consistently creating content.
  • 01:38 — Other Insights from LinkedIn Data: Shows additional data points, including companies he follows grouped by theme, inbound vs. outbound connection requests, and LinkedIn’s perceived interests for ad targeting.
  • 01:58 — The “Game Changer”: Combining Data for Job Search Analysis: He reveals the most impactful feature: using LinkedIn data in conjunction with a job description to perform a gap analysis.
  • 02:05 — Example Job Posting: Director of Information Technology: Presents a real job posting from the Charleston Water System for the AI to analyze.
  • 02:22 — AI Identifying Strengths: The tool analyzes his profile against the job description, highlighting strong qualifications like PhD in Cyber Ops, GRC experience, security architecture, and communication skills.
  • 02:37 — AI Identifying Gaps: The analysis reveals areas where his profile has gaps, such as SCADA/ICS/OT experience (high concern), IT Operations Management, and direct utility/water sector experience.
  • 03:24 — Key Connections to Close Gaps: The “game changer” is unveiled: the AI cross-references his network and identifies specific connections who can help him address the identified knowledge gaps. Cheryl Auger is the top recommendation for SCADA/ICS knowledge.
  • 03:59 — Recommended Action Plan: The AI generates an action plan including who to contact (Rhonda Rummerfield at the hiring org, Cheryl Auger for SCADA insights), resume adjustments, and talking points for the interview.
  • 4:46 — AI for Framing Existing Experience: The tool also provides advice on how to frame his current experience to match the job requirements, emphasizing relevant aspects like NIST 800-53 and HIPAA compliance.
  • 5:20 — How to Perform Your Own Analysis: The speaker outlines the step-by-step process for viewers to replicate this analysis for their own careers.
  • 5:26 — Step 1: Request LinkedIn Data Export: Instructions on how to download a complete archive of personal data from LinkedIn settings.
  • 5:49 — Step 2: Unzip the Data Files: Explains that the export will yield about 50 CSV files with various data points.
  • 5:57 — Step 3: Use Claude Code: Recommends using Claude Code (specifically the Pro version for larger context windows) for the analysis.
  • 6:34 — Inputting Data and Prompts: Demonstrates how to point Claude Code to the unzipped data folder and use specific prompts to analyze LinkedIn data against a job description.
  • 6:52 — Value of AI Analysis and Call to Action: Reiterates the profound impact of this tool for understanding one’s career landscape, finding relevant contacts, and preparing for job opportunities. Encourages viewers to try it and share their findings.
  • 7:16 — Your LinkedIn Network is a Gold Mine: Emphasizes that most people can’t see the value in their raw LinkedIn data, but AI can make it visible and actionable.
  • 7:43 — Final Encouragement to Download Data: A strong call for all viewers to download their LinkedIn data.
  • 8:04 — Find Your “Cheryl Auger”: Inspires viewers to use the tool to identify their own hidden connections who could be pivotal for their career.
  • 8:18 — Resources in Description: Mentions links to Logan’s blog post, Claude Code, and the exact prompts used in the video are available in the video description.

Video 2

URL: https://www.youtube.com/watch?v=EHpAQwXmseQ

Summary

Perplexity Computer represents a significant leap in artificial intelligence, evolving beyond traditional AI search engines to become a fully autonomous work system. Introduced as a cloud-based platform, it orchestrates 19 distinct AI models simultaneously to execute complex tasks. Unlike chatbots that merely answer questions, Perplexity Computer acts as a “digital coworker” – users describe the desired outcome, and the system intelligently plans, delegates tasks to the most suitable AI models, and executes the entire workflow in the background, potentially over weeks or months. The system’s core is powered by Claude Opus 4.6, which acts as the orchestrator, coordinating specialized models like Gemini for deep research, ChatGPT 5.2 for long-context recall and broad web searches, Grok for quick tasks, Nano Banana for image generation, and Veo 3.1 for video generation. This multi-model approach allows for diverse and intricate project completion, from extensive research and report generation to coding applications and creating presentations. A key advantage highlighted is its operation in a secure, cloud-based sandbox environment, ensuring local files on a user’s computer are not directly accessed or impacted. The video showcases several practical demonstrations of Perplexity Computer’s capabilities. These include generating a comprehensive 16-page PDF report on AI in K-12 education, building an interactive web app dashboard comparing top AI tools with features and pricing, creating a four-week YouTube content plan complete with video titles, thumbnails, and key talking points, and even developing a simple web application for tracking monthly subscriptions. Another compelling example is setting up a recurring weekly AI news briefing, which scans and summarizes the latest developments in AI. Currently, Perplexity Computer is exclusively available through the Perplexity Max subscription, priced at $200 per month, which includes 10,000 credits for Computer tasks, with an additional 20,000 bonus credits for new sign-ups (expiring after 30 days). While the cost positions it as a tool for power users and professionals, Perplexity has indicated future plans to extend access to Pro and Enterprise subscribers. Tips for effective use include describing the desired outcome rather than step-by-step instructions, being specific about format and deliverables (e.g., requesting charts or PDFs), utilizing its cross-referencing and analysis strengths, and requesting visual outputs. In conclusion, Perplexity Computer is presented as a cutting-edge AI agent that signifies the future direction of artificial intelligence. By seamlessly integrating and orchestrating multiple specialized AI models, it moves beyond conversational AI to autonomously perform diverse and sophisticated tasks. This capability to handle entire projects with a single prompt offers significant time-saving potential for professionals, marking an exciting evolution towards more capable and integrated AI work systems.

Key Timestamps

Here’s a list of key moments and topics from the video:

  • [00:00] — Introduction to Perplexity Computer, an AI tool that uses 19 different AI models to perform tasks.
  • [00:17] — First glimpse of the Perplexity Computer interface.
  • [00:28] — Overview of Perplexity Computer’s capabilities: research, design, coding, app development, presentations, and project management from a single prompt.
  • [00:55] — The Evolution of Perplexity: From AI Search to Deep Research, Labs, Comet Browser, and now Perplexity Computer.
  • [01:17] — Explaining Perplexity Computer as a “digital co-worker” that breaks down tasks, assigns them to appropriate AI models, and runs everything automatically.
  • [01:35] — Detailed explanation of the 19 AI models Perplexity Computer orchestrates, including Claude Opus 4.6 (Orchestrator), Gemini (Deep Research), ChatGPT 5.2 (Long-Context & Search), Grok (Speed & Quick Tasks), Nano Banana (Image Generation), and Veo 3.1 (Video Generation).
  • [02:23] — Key differences between Perplexity Computer and regular AI chatbots: it plans full workflows, creates sub-agents, runs tasks in parallel, and operates in a secure cloud sandbox.
  • [03:00] — Pricing and access to Perplexity Computer, currently exclusive to the Perplexity Max plan ($200/month) with future plans for Pro and Enterprise users.
  • [03:55] — Walkthrough of the Perplexity Computer user interface, including sections for tasks, files, connectors (for integrating apps like Google Drive, Notion, Slack), and live examples.
  • [04:41] — Beginning of practical demonstrations of Perplexity Computer.
  • [05:02] — Demo 1: Research Task. The presenter prompts Computer to research “the current state of AI in K-12 education.” The AI generates a 16-page PDF report with sections, key findings, and source citations.
  • [06:34] — Demo 2: Interactive Dashboard Creation. The presenter asks Computer to create an interactive dashboard comparing the top 10 AI tools of 2026, including pricing, features, and recommendations. Computer researches, designs, codes, and deploys a live web application.
  • [08:34] — Demo 3: Content Plan Generation. A request to create a 4-week YouTube content plan (titles, thumbnails, key points, tools) for an educational technology channel. Computer produces a printable one-page PDF summary detailing the content plan.
  • [10:25] — Demo 4: Web App Development. The presenter asks Computer to build a simple web app for tracking monthly subscriptions, including service name, price, total monthly cost, yearly projection, and a pie chart. Computer delivers a functional “SubTracker” web app.
  • [12:10] — Demo 5: Weekly AI News Briefing Setup. Computer is instructed to set up a weekly AI news briefing, scanning developments in AI for education and content creation, compiling top 5 stories with summaries and links. Computer sets up a recurring task.
  • [14:15] — Tip 1: Describe the Outcome, Not the Steps – focus on the final result, not how to achieve it.
  • [14:27] — Tip 2: Be Specific About Format & Deliverables – clearly state the desired output format (e.g., PDF, web app, charts).
  • [14:43] — Tip 3: Ask for Cross-Referencing & Analysis – leverage Computer’s ability to compare sources and find disagreements for richer results.
  • [14:57] — Tip 4: Request Visual Outputs – ask for charts, graphs, interactive timelines, or dashboards to utilize specialized models.
  • [15:10] — Tip 5: Keep an Eye on Your Credits – start with smaller tasks to understand credit usage before scaling up to complex projects.
  • [15:20] — Concluding thoughts on Perplexity Computer’s value, particularly for professionals and power users, and its role as a glimpse into the future of AI.

Video 3

URL: https://www.youtube.com/watch?v=09sFAO7pklo

Summary

This video argues that while discussions around Artificial Intelligence often focus on the capabilities of the underlying “models” (like GPT-4 or Claude Opus), the more critical and often overlooked aspect is the “AI harness.” The speaker defines the model as the raw intelligence that understands requests and generates responses, often compared in headlines. In contrast, the “harness” encompasses everything else: where the AI operates (local machine vs. cloud), what it remembers across sessions, what external tools and systems it can access (project management, design files, test systems), how it coordinates tasks, and ultimately, how it fits into and enhances human workflows. The core argument is that models determine how smart an AI is at predicting the next token, but harnesses determine how usefully it integrates into work. The video highlights that while AI models are rapidly converging in their core capabilities, the architectural philosophies of their harnesses are fundamentally diverging. For instance, Anthropic’s Claude Code/Cowork adopts an “incrementalism” approach. It uses a two-part agent system: an initializer to set up projects with structured artifacts (feature lists, progress logs, git history) and a coding agent that makes incremental progress, forcing verification through browser automation. Its philosophy, likened to “bash is all you need,” allows the AI extensive access to the local environment and uses Unix primitives to chain tools. Conversely, OpenAI’s Codex harness, as seen in their internal development, positions the code repository as the “system of record” for everything. It emphasizes a progressive disclosure system for documentation, a rigid layered architecture with linters for enforcement, and runs tasks in isolated cloud containers, managing risk through strict sandboxing and mechanical adherence to defined structures. These diverging philosophies lead to distinct trade-offs. Anthropic’s approach offers flexibility and local integration, but relies more on human oversight and careful artifact management. OpenAI’s method prioritizes safety and mechanical enforcement within a constrained environment, potentially limiting dynamic interaction with local tools unless they are deeply integrated. The speaker emphasizes that this divergence is not just about preference but reflects deeply held theories about how humans and AI should collaborate and where institutional knowledge ought to reside. He cites a study showing the same Claude model performed vastly differently (78% vs. 42%) depending on the harness it ran within, illustrating the profound impact of the harness. Ultimately, the video concludes that for developers and non-technical leaders, the question isn’t which AI model is superior, but which AI harness’s evolutionary trajectory aligns with their team’s long-term workflow and strategic goals. Adopting a harness is a significant strategic commitment, akin to choosing a cloud provider in the early days of cloud computing, impacting operational velocity, security posture, hiring capabilities, and future switching costs. Understanding these technical details of AI harnesses is crucial to avoid expensive lock-ins and effectively leverage AI, shaping how both developers and non-technical knowledge workers will experience the world of work in the coming years.

Key Timestamps

Here’s a breakdown of the key moments and topics in the video:

  • [0:00] — Introduction: AI Harnesses are shaping your working life, but nobody is talking about them enough.
  • [0:15] — Distinguishing the AI “Model” from “Everything Else”: The model is the intelligence (understanding, generating responses), and “everything else” is the harness.
  • [0:28] — Defining the “Harness”: It determines where the AI does its work (local vs. server), its memory across sessions, its access to project tools (management, design, test systems), and its ability to coordinate multiple tasks.
  • [0:55] — The Importance of the Harness: The speaker argues the harness matters “much, much more” than the model because the model only determines how smart the AI is, while the harness determines how it fits into your work.
  • [1:41] — Critiquing Current Comparisons: Most comparisons focus solely on the “model” (likening it to a “brain in a jar”) and ignore the crucial role of the harness.
  • [2:17] — Diverging Harness Architectures: Unlike models, harnesses are diverging rapidly and intentionally, reflecting fundamentally different ideas about human-AI collaboration.
  • [3:05] — Architectural Decisions Lead to Lock-in: Choosing a harness is an architectural decision around which teams build habits, processes, and integrations. Switching later means resetting this entire accumulated value.
  • [3:50] — Models Converge, Harnesses Diverge: Anthropic and OpenAI simultaneously released new coding models. While the models’ capabilities are converging, their harnesses are not; this divergence is the real story.
  • [4:27] — Key Example of Harness Impact: The same Claude model scored 78% on a benchmark when running in Claude Code’s harness, but only 42% when running in a different harness (“Small Agents”), demonstrating the dramatic influence of the harness.
  • [5:08] — Harness as a Performance Multiplier: The harness isn’t just an optimization layer; it’s a performance multiplier that dictates if the model’s intelligence translates into useful work.
  • [5:34] — Two Different Philosophies: The video explores the distinct architectural philosophies of Anthropic’s Claude Code and OpenAI’s Codex.
  • [5:40] — Anthropic’s Harness Philosophy: Built to solve the problem of AI losing context across tasks (one-shotting).
  • [6:09] — Anthropic’s Solution (Claude Code): Uses a two-part pattern (initializer and coding agent) that forces incrementalism, maintains “institutional memory” via structured artifacts (progress files, Git history), and forces verification through browser automation.
  • [7:53] — Claude Code’s Execution Model: Runs in the user’s local terminal with full access to their environment, following a “Bash is all you need” philosophy, chaining Unix primitives for flexibility.
  • [8:31] — OpenAI’s Harness Philosophy: Built around the idea of the repository as the “system of record” for everything (architecture decisions, product principles).
  • [9:08] — OpenAI’s Solution (Codex): Enforces a rigid, layered architecture with documentation, validated dependencies, and linters (some written by Codex itself). Code runs in isolated cloud containers (sandboxes) for security, with internet access disabled by default.
  • [10:13] — Codex’s Execution Model: Runs tasks in isolated cloud containers, using a JSON-RPC harness to programmatically expose and call tools (like Chrome DevTools for UI interaction).
  • [10:43] — Divergence in Memory/State Management: Anthropic makes the agent remember (via structured artifacts), while OpenAI makes the codebase remember (via repo documentation and linters).
  • [10:55] — Practical Differences for Users (Calvin French-Owen’s experience): He uses Claude Code for planning and orchestration (where flexibility and local access are key) and Codex for implementation (where code quality and isolation are paramount).
  • [12:24] — Divergence 1: Execution Philosophy (Deep Dive): Anthropic prioritizes composability and local environment access, while OpenAI focuses on controlled, observable execution in the cloud for UI/application interaction.
  • [14:45] — Divergence 2: State and Memory (Deep Dive): Anthropic emphasizes explicit, structured artifacts for agent memory, while OpenAI relies on encoding knowledge directly into the codebase and using background processes to maintain architectural integrity.
  • [16:36] — Divergence 3: Context Management (Deep Dive): Anthropic compacts context and delegates to sub-agents, loading skills just-in-time. OpenAI emphasizes isolation and controlled access to relevant information in its sandboxed environments.
  • [17:56] — Divergence 4: Tool Integration (Deep Dive): Both use the Model Context Protocol (MCP), but Anthropic’s integration is more flexible (Unix primitives), while OpenAI’s is deeper and assumes a server-mediated cloud environment.
  • [19:56] — Divergence 5: Multi-Agent Architecture (Deep Dive): Anthropic uses an orchestrated collaboration model with multiple sub-agents, shared task lists, and human oversight. OpenAI’s approach is more isolated, with tasks running in separate sandboxes, relying on code commits for coordination.
  • [21:46] — Strategic Importance: The question shifts from “which model is best?” (a short-lived, tactical question) to “which harness’s evolutionary trajectory matches our team’s direction?” This is a strategic commitment, impacting velocity, security, hiring, and switching costs.
  • [22:54] — Understanding Harnesses for All: The technical details of harnesses are crucial not just for engineers, but for non-technical knowledge workers and leadership, as these architectures will shape how everyone experiences work in the coming years.

Graph View