So What Growing Prompts into AI Skills - CourtneyR.dev's 2nd Brain

So What? Growing Prompts into AI Skills 🔗 **Source:** [So What? Growing Prompts into AI Skills](https://www.youtube.com/watch?v=s8Qm-0iFZLQ) > [!note]+ Document Notes > tags: #artificial-intelligence #prompt-engineering #automation #workflow-optimization #skill-development > author: Trust Insights > source: YouTube > reference: https://www.youtube.com/watch?v=s8Qm-0iFZLQ > date created: 2024-12-19 14:30 > > # 🏷️ Title: So What? Growing Prompts into AI Skills >  ## 🧰 Action Inbox - [H1] Build prompt-to-skill conversion system — 06:45 #automation - [H2] Create centralized skill repository using GitHub — 45:30 #workflow-optimization - [H3] Test all skills with validation scenarios — 25:15 #quality-assurance ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [Agent Skills Documentation](https://agentskills.io) - [Anthropic Skills Guide](https://anthropic.com/skills-guide) ## 📝 Summary - **What happened**: Trust Insights demonstrates evolving AI prompts into portable "Skills" that work across multiple AI platforms - **Why it matters**: Skills eliminate the platform lock-in problem of custom GPTs and Gems while enabling automation at scale - **Key stakeholders**: Marketers, developers, teams with existing prompt libraries - **Risks and tradeoffs**: Requires structured approach and testing; skills can malfunction if poorly designed - **Bottom line**: Skills represent the next evolution from individual prompts to automated, portable AI workflows ## 📖 Main Content Skills are Anthropic's open-source format for creating portable AI workflows that function like custom GPTs but work across any compatible platform. Unlike platform-specific tools, Skills use standardized YAML formatting and can be moved between Claude, Google Anti-gravity, OpenAI Codex, and other systems. The presentation outlines a five-level AI maturity model: Level 1 (manual prompting), Level 2 (custom GPTs/Skills), Level 3 (project planning), Level 4 (autonomous agents), and Level 5 (agent swarms). Skills primarily operate at Level 2 but enable progression to higher levels. The Trust Insights team demonstrates their custom prompt-to-skill conversion tool, which applies their 5P framework (Purpose, People, Platform, Process, Performance) to transform existing prompts into properly structured Skills with built-in testing and validation. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/s8Qm-0iFZLQ?t=0) — [H1] Introduction to AI Skills Evolution - Key idea: AI prompt usage has evolved from basic frameworks to automated, portable Skills - Evidence: "Three plus years ago when generative AI hit the public market, everyone was struggling to figure out how to prompt correctly" #### ⏱️ [02:15](https://youtu.be/s8Qm-0iFZLQ?t=135) — [H2] Skills vs Custom GPTs Comparison - Key idea: Skills solve the platform lock-in problem of custom GPTs and Gems - Evidence: "A skill is portable. It is effectively a prompt or set of prompts that is bundled together and you can move it to any system that supports skills" #### ⏱️ [04:30](https://youtu.be/s8Qm-0iFZLQ?t=270) — [H1] Five-Level AI Maturity Model - Key idea: Organizations progress through defined stages of AI adoption and capability - Evidence: "Level one: done by you. Level two: done with you. Level three: project planning. Level four: autonomous agents. Level five: organizational division" #### ⏱️ [08:45](https://youtu.be/s8Qm-0iFZLQ?t=525) — [H2] Skills Positioning in Maturity Model - Key idea: Skills operate primarily at Level 2 as Standard Operating Procedures - Evidence: "Skills fall in at level two. Skills are essentially SOPs" #### ⏱️ [12:00](https://youtu.be/s8Qm-0iFZLQ?t=720) — [H1] Five P Framework for Skill Building - Key idea: Structured approach ensures comprehensive skill development and testing - Evidence: "Purpose, People, Platform, Process, Performance - a well structured SOP covers all of these pieces" #### ⏱️ [15:30](https://youtu.be/s8Qm-0iFZLQ?t=930) — [H2] Skill Triggering and Evaluation - Key idea: Proper skill design includes when to fire and validation scenarios - Evidence: "If it's badly written, your skill will fire at the most inopportune times and screw everything up" #### ⏱️ [20:15](https://youtu.be/s8Qm-0iFZLQ?t=1215) — [H3] Software Development Lifecycle Application - Key idea: Building skills requires following software development best practices - Evidence: "You're developing software you just don't know it. The software development life cycle: requirements, design, develop, test, deploy" #### ⏱️ [25:15](https://youtu.be/s8Qm-0iFZLQ?t=1515) — [H1] Testing and Validation Requirements - Key idea: All skills must include built-in testing to prevent failures - Evidence: "Do not leave it up to chance. Make sure there is some kind of test, some sort of validation at the end of every skill" #### ⏱️ [30:45](https://youtu.be/s8Qm-0iFZLQ?t=1845) — [H2] Live Demonstration of Skill Creation - Key idea: Converting existing prompts to skills can be automated using structured tools - Evidence: Demonstrates converting a Suno music generation prompt into a portable skill #### ⏱️ [35:30](https://youtu.be/s8Qm-0iFZLQ?t=2130) — [H3] Anti-Pattern Recognition - Key idea: Skills must define what NOT to do, not just positive instructions - Evidence: "Anti-patterns is the opposite of a design pattern. These are the critical mistakes that the system needs to be told don't do this" #### ⏱️ [40:15](https://youtu.be/s8Qm-0iFZLQ?t=2415) — [H1] Mass Conversion Strategy - Key idea: Convert entire prompt libraries at once rather than individual prompts - Evidence: "You do not convert those prompts one at a time. Put the prompts as text files in a folder and process every one with its own agent" #### ⏱️ [43:00](https://youtu.be/s8Qm-0iFZLQ?t=2580) — [H2] Command Line Integration Capabilities - Key idea: Skills can integrate with enterprise systems through CLI tools - Evidence: "Google workspace has a command line tool that your AI can use to reach into Google Keep and pull out your company prompt library" #### ⏱️ [45:30](https://youtu.be/s8Qm-0iFZLQ?t=2730) — [H1] Centralized Skill Management - Key idea: Use GitHub repositories for company-wide skill distribution and updates - Evidence: "Have a private company GitHub repository as your skills marketplace. When you make changes, all instances auto update" #### ⏱️ [48:15](https://youtu.be/s8Qm-0iFZLQ?t=2895) — [H3] Automated Scheduling Capabilities - Key idea: Skill conversion and management can be fully automated on schedules - Evidence: "You could have a monthly rescheduled task that automatically converts and publishes and runs the whole workflow" ## 📚 References - [Agent Skills Specification](https://agentskills.io) - [Anthropic Claude Skills Guide](https://anthropic.com) - [Trust Insights Prompt Playbook](https://trustinsights.ai/prompt-playbook) - [Trust Insights 5P Framework](https://trustinsights.ai/5p-framework) ## 🗣️ Paraphrase - Skills are portable AI workflows that solve the platform lock-in problem of custom GPTs by using open standards - Organizations progress through five levels of AI maturity, from manual prompting to autonomous agent swarms - The 5P framework (Purpose, People, Platform, Process, Performance) provides structure for converting prompts into robust skills - Proper skill development requires software development practices including testing, validation, and anti-pattern definition ## 📰 Newsletter **Transform Your Prompt Library into Portable AI Skills** Ready to break free from platform lock-in? Skills are the next evolution of AI automation, letting you build once and deploy anywhere. Unlike custom GPTs that only work in ChatGPT, Skills use open standards and work across Claude, Gemini, and other platforms. **Key Benefits:** - Portable across AI platforms - Built-in testing and validation - Scalable automation capabilities - Enterprise-ready with CLI integration **Getting Started:** Apply the 5P framework to your existing prompts: Purpose, People, Platform, Process, Performance. This structured approach ensures your skills are robust and reliable. **Pro Tip:** Don't convert prompts one at a time. Use tools like Claude Code to batch-process your entire prompt library overnight. Resources: [Agent Skills Documentation](https://agentskills.io), [5P Framework](https://trustinsights.ai/5p-framework), [Prompt Playbook](https://trustinsights.ai/prompt-playbook) ## ✅ Action Items #### DevRel - Developer advocacy: Create Skills showcase demonstrating cross-platform compatibility - Education: Develop Skills workshop curriculum for developer communities - Developer experience: Build Skills validation tools and testing frameworks - Developer marketing: Launch "Skills vs GPTs" comparison campaign - Community management: Moderate Skills sharing forums and gather feedback #### OSPO - Policy & Governance: Establish Skills contribution and review policies - License Compliance: Review Anthropic's open source Skills format licensing - Security Management: Create Skills vulnerability scanning and review processes - Software Inventory: Track Skills dependencies and maintain SBOM documentation - Developer Support: Provide Skills development training and tooling consultation - Community Engagement: Fund upstream Skills specification development - Strategic Alignment: Measure Skills adoption rates and developer productivity gains #### Community - Maintainers: Support Skills format adoption across open source projects - Contributors: Create first-issue programs for Skills development - Sponsors: Fund Skills marketplace infrastructure and documentation - Ecosystem: Coordinate cross-platform Skills compatibility testing #### WordPress - Core: Evaluate Skills integration for WordPress development workflows - Core Performance: Build Skills for performance profiling and optimization tasks - Core AI: Develop Skills for AI-assisted code review and testing - Community: Create Skills for meetup planning and WordCamp coordination - Training: Convert existing lesson plans into interactive Skills format - Documentation: Transform documentation workflows into reusable Skills - CLI: Build WP-CLI Skills for common administrative tasks #### The WPCC - Programs: Develop Skills for fellowship application processing and review - Projects: Create governance template Skills for open source projects - Funding: Build transparent budget tracking and reporting Skills #### FAIR - Technical independence: Ensure Skills format remains vendor-neutral - Policies: Create Skills marketplace governance and quality standards - Governance artifacts: Document Skills specification evolution and decision-making ## 🧪 Hypotheses - Organizations using Skills will achieve 50% faster AI workflow deployment compared to platform-specific tools - Teams with centralized Skills repositories will show 30% higher AI adoption rates - Skills built with the 5P framework will have 75% fewer runtime failures than ad-hoc conversions - Companies implementing Skills automation will reduce prompt maintenance overhead by 60% ### 📊 Methodology - Study design: Compare deployment times between Skills and custom GPTs across multiple organizations - Data sources: Track Skills repository usage metrics and error rates from GitHub analytics - Experiments: A/B test 5P framework vs. unstructured prompt-to-skill conversions with failure rate analysis ### ❓ Questions - How will Skills format evolution affect backward compatibility across platforms? - What governance models work best for enterprise Skills marketplaces? - Can Skills format support real-time collaborative editing and version control? ### 🔮 Future Research - Integration patterns between Skills and enterprise workflow automation platforms - Skills marketplace monetization models and sustainability approaches - Cross-platform Skills testing automation and continuous integration pipelines ### 💡 Implications - Practice: Teams need Skills development training and centralized repository management - Policy: Organizations require Skills governance frameworks and security review processes - Security: Skills must be scanned for vulnerabilities and malicious code injection - Education: Developer curriculum should include Skills format and automation best practices - Funding: Budget allocation needed for Skills infrastructure and marketplace development ### 🧠 Learning Styles Assessment Reading level estimate: 12th grade (technical content with specialized terminology) #### 👀 Visual learners - Present: Five-level maturity model diagram, skill architecture screenshots - Improve: Skills workflow flowcharts, before/after prompt transformation visuals #### 🔊 Aural learners - Present: Live demonstration with verbal explanation, discussion format - Improve: Audio-only Skills tutorial series, podcast deep-dive episodes #### 🗣️ Verbal learners - Present: Framework definitions, detailed explanations, analogies to software development - Improve: Skills vocabulary glossary, comparison tables with other automation tools #### 🧩 Kinesthetic learners - Present: Live coding demonstration, hands-on skill building process - Improve: Interactive Skills builder sandbox, step-by-step conversion workshops #### 🧠 Logical learners - Present: Five-level progression model, structured 5P framework, systematic approach - Improve: Skills ROI calculators, decision trees for when to use different approaches #### 🧑‍🤝‍🧑 Social learners - Present: Team discussion format, community marketplace concept - Improve: Skills sharing forums, collaborative building sessions, peer review processes #### 🧘‍♀️ Solitary learners - Present: Self-guided skill building process, individual prompt library conversion - Improve: Personal Skills audit worksheets, self-assessment rubrics #### 🌿 Naturalistic learners - Present: Evolution metaphor from prompts to skills - Improve: Ecosystem thinking around Skills marketplaces, organic growth patterns ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive link text for resources - Improve: Add ARIA labels for live demonstration screenshots, alt text for diagrams #### Keyboard navigation - Present: Standard video player controls accessible via keyboard - Improve: Transcript navigation shortcuts, skills documentation keyboard shortcuts #### Contrast and readability - Present: High contrast video interface, clear typography - Improve: Ensure code examples meet contrast requirements, larger font options #### Text resizing and reflow - Present: Responsive video player, scalable interface elements - Improve: Ensure skills documentation supports 200% zoom without horizontal scrolling #### Media and captions - Present: Full transcript provided for accessibility - Improve: Add chapter markers for easier navigation, downloadable SRT files #### Forms and interactions - Present: Simple poll interface with clear labels - Improve: Keyboard-accessible skills builder interfaces, error handling for form validation #### Standards reference - Map video accessibility to WCAG 2.2 AA guidelines for multimedia content ### 🌍 DEIB Assessment #### Cultural breadth - Present: Generic examples applicable across industries and regions - Improve: Include Skills examples from diverse cultural contexts, multilingual documentation #### Inclusive language - Present: Gender-neutral pronouns, avoided technical jargon explanations - Improve: Plain language summaries for complex concepts, cultural context for business examples #### Neurodiversity support - Present: Structured presentation with clear sections, consistent formatting - Improve: Executive summary sections, visual progress indicators for multi-step processes #### Representation - Present: Mixed team presentation with diverse perspectives - Improve: Include Skills examples from various industries, highlight diverse use cases #### Psychological safety - Present: Encouraging tone about learning and experimentation - Improve: Normalize mistakes in Skills development, provide supportive troubleshooting guidance ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear progression from basic prompts to automated skills --- ## Full Transcript [00:12] Heat. Heat. [00:33] Well, hey everyone. Happy Thursday. [00:34] Welcome to So What the Marketing [00:36] Analytics and Insights live show. I am [00:38] Katie joined by Chris and John. [00:40] >> Hello fellows. [00:42] >> There you go. [00:45] Uh this week we are talking about [00:48] growing prompts into AI skills. So [00:51] threeish plus years ago when generative [00:54] AI hit the public market for the general [00:58] consumer, everyone was struggling to try [00:59] to figure out how to prompt correctly, [01:02] prompt effectively, prompt efficiently. [01:04] So that's where things like the trust [01:07] insights repel framework were born. Uh [01:10] there's also uh we we've had image [01:13] prompts, we've had race, we've had pair. [01:16] I mean, there's a lot of different [01:18] frameworks for um I don't I lost I don't [01:22] know how to shut this one off. There we [01:23] go. Um there's a lot of different [01:26] frameworks for how to do effective [01:27] prompts. However, now uh you have [01:31] systems that can build in skills and [01:36] take all of your prompts and you know [01:38] automate them. So sort of the the [01:40] transition was so you have a general [01:42] prompt structure, you paste that in [01:44] every time. Then you build a custom GPT [01:47] or a gem or a project depending on the [01:50] system that you're in and those have [01:52] your prompts baked in so you don't have [01:54] to copy and paste them over and over [01:56] again. So now the next evolution of that [01:58] is how do you take that custom thing and [02:01] make it a skill that you can apply [02:05] across anything that you're doing so you [02:07] don't just have to work within that [02:08] custom gem or that custom GPT or that [02:11] project. And that's what Chris is going [02:14] to walk us through today. And by us I [02:17] mean me and John and then whoever's [02:20] watching. [02:24] So it let's start with a bit of table [02:26] setting on skills. Skills is a specific [02:28] term skills with a capital S. This is a [02:31] data format that anthropic uh the makers [02:34] of Claude originally came up with for [02:37] Claude its ecosystem. Uh it is the best [02:41] way to think about it is kind of like a [02:43] GPT or a gem that you can move anywhere. [02:47] So one of the biggest criticisms of [02:48] things like custom GPTs was okay you [02:50] have to use chat GPT with it right or if [02:52] you have a gem you have to use Gemini [02:53] with it. A skill is portable. It is [02:56] effectively a prompt or a set of prompts [02:59] that is bundled together and then you [03:01] can move it to any system that supports [03:03] skills and then anthropic made it a [03:05] public good. They open sourced it and [03:07] they said anyone can use the skills [03:09] format. Uh and in fact there is a whole [03:12] website uh called agentskills.io IO that [03:16] is Anthropic's official this is the [03:18] definition of the skills format. So this [03:20] is how you use it and there's um they [03:22] have this they have like a 32page slide [03:26] deck that explains what goes into making [03:28] a good skill and things like that. [03:31] skills are supported by Claude Code, [03:34] Claude Co-work, Google Anti-gravity, [03:37] OpenAI Codeex, Quen Code, Open Cowwork, [03:41] uh Claude Cowwork, uh and pretty much [03:44] most agentic frameworks now because [03:47] they're so dang useful. [03:50] One of the things that I want to take a [03:52] step back on is a skill is because it's [03:56] essentially a GPT that you can move [03:58] anywhere. uh it fits in kind of in an a [04:02] specific place in the evolution of AI. [04:04] So I want to talk about this briefly. [04:07] This is something that at Trust Insights [04:09] we've been working on for for quite some [04:10] time. Um originally derived from product [04:12] market fit. It's now sort of evolved [04:14] onto its own thing. And there's five [04:16] levels. There's done by you which is [04:18] level one. That's you and chat GPT and [04:20] you are the typing monkey. And you're [04:22] just chatting how do I do this? What's [04:24] the weather? Where's a good place to go [04:25] eat? you know, should I vacation in uh [04:28] Montenegro, etc. And that's fine. You're [04:30] doing all the work. A lot of value [04:32] there. Learn how to do prompting, right? [04:34] Uh we've talked about that for, as Katie [04:36] said, three years. Second level done [04:39] with you. These are gems in GPTs. [04:43] Effectively, what you're doing is you're [04:45] building in a standard operating [04:47] procedure into into one of these things. [04:50] So that you have your uh '9s uh c uh [04:53] song lyrics generator as an example of a [04:56] of a standard operating procedure at the [04:59] agent level. Now you're doing project [05:02] planning. So I think it was on the [05:05] podcast we talked about this where we [05:07] said you know the hottest programming [05:09] language of 2026 is project management [05:11] if you're good at project management. [05:14] So that is project plans. That's what [05:17] you get you do get there. Today you now [05:21] have fully autonomous or mostly [05:23] autonomous systems like OpenClaw uh and [05:26] the brand new as of yesterday claude [05:28] managed agents uh that are built right [05:31] into the cloud ecosystem. So what do you [05:34] suppose that [05:37] uh input is for that [05:40] >> for [05:42] the managed agents? [05:43] >> Yeah. Oh. Um, I would assume all of your [05:46] skills and plugins [05:50] >> and your project plans. [05:51] >> Yep. Well, and so and as I'm saying it, [05:54] I realize like these each build on each [05:56] other. So, in phase one, you start with [05:59] the prompts. In phase two, you have [06:01] prompts plus. In phase three, you have [06:03] plus and plus. And so, you don't just do [06:06] something brand new in each of these [06:08] levels. You build on what you did in the [06:11] previous level. So, you still have to [06:13] master level one, level two, level [06:16] three. I would say I'm squarely in level [06:20] three with like a little bit of dabbling [06:22] in level four. [06:24] >> You're going to you're going to call [06:26] yourself wrong in just a moment. Um [06:29] because you you already have the skills, [06:32] Katie, um to do levels four and five. Um [06:36] you may not know it, but you do. And [06:40] here's why. [06:42] At level four, when you're building [06:44] something like an open clock, you're [06:45] providing a job description. A job [06:47] description is a bundle of project plans [06:49] plus skills plus plugins, plus all that [06:51] stuff. Yet, you say, "Agent, you're [06:53] you're my stock broker. Go off and do [06:55] it." Right? So, it effectively is that [06:57] point is like a full employee that you [06:59] just go off. [07:00] >> Level five is an organizational [07:03] division. to the strategic charter where [07:05] you would have a a swarm of agents with [07:09] a supervisor agent running them. And so [07:12] someone like you, Katie, who has built [07:14] multi-million dollar teams and managed [07:17] teams of people, you have all the skills [07:19] you need to be able to create a [07:20] strategic charter for an organizational [07:22] unit like a a division of a company [07:25] instead of just a a a single employee. [07:27] >> And you could send that division off and [07:29] say, "This is my marketing division. [07:31] I've got an SEO agent. I've got a PPC [07:33] agent. I've got an email agent. I've got [07:35] an a marketing manager agent. And you as [07:37] that strategic charter would provide all [07:39] the job descriptions and all the skills [07:41] and all the plugins and things like [07:42] that. And this the agent swarm goes off [07:45] and does those things and comes back to [07:47] you once a month and says, "Hey Katie, [07:49] the u the uh virtual marketing [07:51] department did all these things for [07:53] you." [07:54] Hm. [07:56] Having worked with humans, my guess is [07:59] that, you know, if they only check in [08:01] once a month and I look at the report, [08:02] I'm like, "That's not what I asked you [08:04] to do. Why didn't you do the things I [08:05] wanted you to do? What the heck? You [08:08] spent all this money." Like, no. But, I [08:10] mean, I can definitely see the potential [08:11] in that. And so, with that extra [08:13] description that I would say, yeah, I'm [08:15] I'm not squarely in number three. I'm [08:17] squarely in number four at the moment as [08:19] of today. But I can certainly see your [08:22] point about uh getting into number five. [08:26] As a as a human, I'm not [08:31] totally comfortable [08:33] being hands off like that. But you know, [08:36] that's a that's a me problem. [08:40] >> So with that understanding then, where [08:42] do skills fall in? Skills fall in at [08:44] level two. Skills are essentially SOPs, [08:47] >> right? And you can take any any prompter [08:50] conversation or raw materials from level [08:52] one and bring it up into a system and [08:55] turn it to level two. And here's the [08:57] thing, there's so many tools that do [08:59] this right now. If you use, for example, [09:01] Claude Code, Claude Code has a a plug-in [09:04] builder and a skill builder. Uh Claude [09:06] Co has a plug-in builder and a skill [09:08] builder. And the ones that come in the [09:10] package are pretty good. They're not [09:14] great. And so that's what we wanted to [09:16] talk through today is how I built one [09:18] for myself um that we actually put in [09:21] the brand new uh trust insights prompt [09:24] playbook which is a monthly uh monthly [09:26] subscription but the the first lesson is [09:28] literally the plug-in um because [09:33] the existing ones operate on a a [09:36] presumption that what you're putting in [09:38] was good to begin with. [09:41] That's not always the case. Well, the [09:44] other thing too, so like these these are [09:46] what the out of the box skills that you [09:48] can So [09:53] the the they're good to start with, but [09:56] they're generic. So there's no [09:58] customization. There's no tailoring [10:00] towards what you as a human do or what [10:03] your team does or what your organization [10:04] does. So I would, you know, and we've [10:07] talked about this on previous live [10:08] streams. So, like John can use the legal [10:12] plugin or the sales plugin, but he's [10:15] going to have to customize a lot of the [10:18] instructions based on, you know, the [10:21] kinds of terms that we look for in [10:23] contracts or the things that need to be [10:25] called out, you know, in an MSA or, you [10:28] know, the way that we do our sales [10:30] cycle, which there's a lot of [10:32] standardization, but our clients don't [10:34] look the same as other clients. So we [10:36] need to sort of work with them and not [10:37] force them into a certain process. So [10:40] there, you know, for John, like the [10:42] sales and the legal skills are good [10:45] starts, but they're not going to get him [10:47] all of the way into just doing it for [10:50] me. There's still so much work. So, you [10:52] know, for someone like him, it's [10:54] probably easier just to build his own [10:56] skill for the trust insight sales [10:58] process, for the trust insights legal [11:01] documentation review versus using what's [11:03] out of the box. That's probably true for [11:05] most people. [11:07] >> Um, [11:07] >> exactly. [11:08] >> Yeah. [11:09] >> So, how do we do this? The the [11:13] completely uh nonsurprising answer is [11:16] that if you want a good standard [11:18] operating procedure, you should have [11:20] some kind of framework that governs that [11:22] procedure. Um, now we obs [11:27] >> Go ahead. Go ahead. [11:28] >> Can I can I guess? [11:33] >> Sorry. Go ahead. [11:35] >> Well, no. So, it is the 5P framework and [11:38] that is the basis of the prompt to skill [11:41] plugin that we built which uh which [11:43] basically takes the existing prompt you [11:45] have and puts it through the fiveps. And [11:49] what you end up with when you run it is [11:52] you end up with a skill that gets [11:54] autoinstalled into your your clawed [11:56] environment. But you also get an [11:58] understanding and this is the part that [12:00] is missing in most skill builders. you [12:02] don't. It has an eval section that looks [12:05] at whether or not the skill should fire [12:09] or not. And this is the this is the big [12:11] thing. If you've been using skills for a [12:12] hot minute, you know that [12:16] if you have similarly named skills, the [12:19] triggering section, which is the YAML [12:21] front matter of the skill, is what [12:23] determines when the skill runs or not. [12:24] And if it's badly written, [12:28] uh, your skill will fire at the most [12:29] inopportune times and screw everything [12:31] up. Um, so part of of good skill [12:34] building is that. So the first thing you [12:36] need to do is take your prompt and put [12:39] it through the five Ps. The thing you [12:42] want to turn into a skill. Why are we [12:44] doing this? What are we doing? Uh, you [12:47] know, who's going to do it? Who's [12:48] impacted by it? How do you do it? Do you [12:52] need any supplementary tools? any things [12:54] like MCPs or web searches and things and [12:57] then the critical part that is not in [12:59] any of the other skill builders is the [13:01] performance the eval to say does this [13:03] skill work as intended [13:06] a well ststructured SOP covers all of [13:10] these pieces um so if you have well [13:13] ststructured SOPs then you're already in [13:15] good shape if you don't um so I always [13:19] you know tell the story uh the first [13:21] time I walked into the agency to manage [13:23] the team that Chris had hired me to [13:25] manage. The first question I said was, [13:27] "All right, let's take a look at the [13:28] SOPs." And they all look at me like I [13:30] had eight heads and I was like, [13:32] "Standard operating procedures, uh, [13:34] instructions for how you do things." Oh, [13:36] yeah. No, we don't have that. And I was [13:37] like, "Okay, that's number one thing [13:39] we're going to start." And you know, so, [13:42] uh, as Chris mentioned at the top sort [13:44] of with the strategic charter, like [13:47] there's no real secret to building teams [13:51] and growing teams. A lot of it is just [13:53] good structure and SOPs. And those are [13:56] the things that I'm really really [13:57] skilled at. So therefore, having good [14:00] SOPs is going to help you move more [14:02] quickly through these five levels. [14:06] >> Exactly. So with that in mind, let me [14:09] see if I can get the prompt to skill [14:12] skill installed in my Claude co-work. [14:15] Oh, that did not work. So I'm just going [14:16] to use cloud code. Um, let's walk [14:19] through what happens in this thing [14:21] because I think it's a useful exercise. [14:24] So, in [14:26] our skill, let's go look at prompt of [14:28] skill. Make this just a bit bigger so we [14:30] can see it. [14:32] Every skill at a minimum has to have a [14:34] skill.md file and a folder that it's in [14:37] it. And every skill that you build, um, [14:40] and again, the skill builders will do [14:42] this for you to a degree. U will tell [14:44] you, let's make this bigger so that [14:46] we're not straining our eyes. give [14:47] ourselves migraines. [14:50] >> Um the description tells you when to [14:54] fire when should this this thing uh [14:56] work. So this one is convert any [14:58] existing prompt workflow or set of [14:59] instructions to a proper structured [15:00] cloud skill with the skill MD scripts [15:02] references and assets. This is another [15:04] part that again I feel like some of the [15:08] the skill builders out there kind of [15:10] miss. It doesn't have to just be prompts [15:12] inside of a skill. You can have things [15:14] like Python scripts. Now, obviously, if [15:16] you're going to be running Python [15:18] scripts, be 100% sure you trust the [15:21] where the the script came from because [15:23] there's obviously a lot of bad players [15:26] out there. Um, [15:28] and then you have the invocations and [15:31] then trigger and ask to make a [15:32] persistent thing attempts to do these [15:34] things. What happens in this flow? You [15:37] get the extraction, you run the five Ps [15:41] and say, what is this prompt trying to [15:43] do? um why, you know, why would the user [15:46] do this? Who's doing it? Who's impacted [15:48] by it? Um what is the step-by-step [15:50] process? What is what tools, data [15:52] sources, MCPS, and things does it need? [15:54] And then the performance, how do you [15:56] know that it works? It has to write down [15:58] its analysis first because we never ever [16:00] let any agent system do work in memory [16:03] because the moment you quit that [16:04] session, it's gone. So, it always has to [16:06] take notes. Then determine the skill [16:09] category. [16:10] Build the skill name. Write the front [16:13] matter which has the triggering [16:14] capabilities. Build the skill body with [16:18] the corrected structure. Identify [16:20] include the bundled resources. [16:22] This is now the fifth P. The five Ps. [16:25] Generate test cases. So it has to when [16:28] you're doing a skill building, you have [16:29] to test your skill. And if you build as [16:31] part of a skilled builder, you can run [16:34] test cases and then and then look at the [16:36] evals inside [16:39] the plugin [16:41] uh is the eval. So this is an example of [16:45] does the tool know when to fire? And so [16:48] there's five different scenarios that [16:50] contestants say should this skill fire. [16:52] Someone says what's the capital of [16:54] France? The skill should not fire. [16:56] Right? So that's a clear eval case of [16:58] evaluating saying yeah this skill if the [17:00] skill would fire [17:02] something's gone wrong and so that's the [17:06] process of we baked this thing for [17:09] ourselves and it's in the prompt [17:11] playbook um so that you don't have to [17:14] walk through and do this by yourself but [17:16] that is the overall process you need a [17:17] framework you need to take the prompt [17:21] rebuild it through that framework and [17:23] the magic phrase is that I would use if [17:27] you wanted to build this on your own is [17:28] upstream and downstream. [17:31] No prompt exists in a vacuum. It has to [17:34] be in the context of something else that [17:36] you're doing. So a critical question to [17:38] ask and it's built into this the skill [17:40] builder is what are the upstream and [17:42] downstream things around this prompt and [17:44] and these tools particularly a model [17:46] like opus is smart enough to go well you [17:50] know John just gave me a uh Mtown lyrics [17:53] prompt uh to to do some to infer some [17:57] song building. So upstream, uh, I need [17:59] knowledge of what good Mottown songs [18:01] are. And downstream, I'm going to guess [18:03] that John's probably going to use this [18:05] in a tool like Sunno to generate, uh, a [18:08] a song, which means that in addition to [18:11] the lyrics, I should probably create a a [18:14] music composition prompt to go with it. [18:19] >> So what you're describing is the basics [18:22] of software development as well. um you [18:25] know and I think that that's something [18:27] that you know a lot of people stayed [18:29] away from generative AX like I'm not a [18:31] developer I can't code so now you know [18:35] we the general we were calling these [18:37] things oh you're not coding it's a skill [18:40] you're building a plugin you're not [18:42] coding you're not developing software [18:44] guess what you're developing software [18:47] you just don't know it um and so you [18:50] know we talk about the 5P framework as [18:53] the good framework for structuring the [18:55] information. But really the other [18:58] framework that you need to be leaning on [19:01] is the software development life cycle. [19:04] We didn't create it, we just use it. So [19:06] the software development life cycle, the [19:08] SDLC, you know, you start with your [19:10] requirements. That's where the 5P [19:12] framework comes in. Then you have [19:15] design, develop, [19:18] you know, test, deploy, and then do it [19:20] all over again. And those are the basic [19:22] um [19:24] the basic stages of the software [19:26] developer [19:35] and you have those triggers, that's [19:37] that's you testing the software that [19:39] you've built. Um if you are not testing [19:43] things as you're building it, you're [19:46] doing it wrong. There's no two ways [19:47] about it. You have to be testing as you [19:49] go. Do not leave it up to chance. Even [19:52] if you're using systems like Claude Code [19:54] or co-work, they can still hallucinate. [19:58] They can still bring in things you [19:59] didn't mean them to bring in. They can [20:01] still make mistakes. You can make [20:02] mistakes when you're writing the skill. [20:04] Make sure there is some kind of [20:08] test, some sort of validation, call it [20:10] whatever you want at the end of every [20:13] skill to make sure it's doing what it's [20:15] expected. especially as you know so you [20:19] might build a skill today and then 3 [20:20] months down the road you're like hey I'm [20:22] doing this project let me pull in this [20:23] skill and it's a mismatch but you might [20:26] just because of what you've called it [20:28] you think it's working but if there's no [20:30] testing or validation built in you could [20:32] be basically taking all of your results [20:34] and going sideways with it. So test [20:37] everything be disciplined and strict [20:39] about that. [20:41] >> Yep. Exactly. So let's put this to work. [20:45] Let's uh do this. One of the reasons, by [20:47] the way, that I keep mentioning the [20:48] prompt playbook in addition to being a [20:49] narcissistic greeting maniac is that all [20:53] of the prompts that are in it, you can [20:54] drop into the plugin and generate the [20:58] skills. So, if you'd like a prompt for [21:00] because it comes out every Tuesday and [21:01] Friday, if you like a prompt from that [21:02] week, like, oh, this means this was a [21:04] skill. Now, just drop it in say, turn [21:06] this into a skill. Um, I'm gonna pick [21:08] something that is uh not under NDA. Uh, [21:11] and I'm gonna pick [21:12] >> We appreciate that. [21:13] >> Yes. our prompt, our our old gem from [21:17] our Google gem uh gems pile of [21:20] generating [21:22] um music uh stuff specifically for [21:25] Sunno, the music composition tool. So, [21:26] I've got a a a gem in set of [21:29] instructions here. I'm going to copy and [21:31] paste this and right inside of Claude, [21:33] we're going to say, [21:35] let's turn this [21:39] music generation prompt into a skill. [21:45] Uh, and then I'll put it I'll just put [21:47] in three hashes there. Oops. [21:54] And we'll put in our prompt from Gemini. [21:58] So, if you've got chatgptt gems and the [22:01] instructions laying around, you got [22:02] Google gems and things like that, um, [22:04] you can get this to do a really good [22:05] job. If you're in clawed code or a [22:09] similar style of system, anti-gravity [22:11] codeex, etc., go with plan mode first [22:15] always for anything like this. Um, and [22:17] if you're using claude code, go to use [22:20] the opus plan model which does planning [22:23] in opus and then does implementation and [22:25] son. It saves you a lot of money which [22:26] goes back to last week's uh live stream [22:28] about how do we um how do we manage AI [22:30] limits. So, this is going to fire up the [22:33] prompt to skill skill because it [22:35] correctly identified I turn I said I [22:36] want to turn this prompt into a skill. [22:38] And now it's going to look through and [22:40] try and figure out how to do this. [22:43] When it's done, what it will do is it [22:46] will say, I've I've essentially figured [22:48] out this skill and now I'm going to [22:50] build it for you and I'm going to [22:51] install it for you. U so you really have [22:54] to do very very little beyond that. [22:56] But again, it sort of reinforces making [22:59] sure that your plan upfront [23:02] is as thorough and detailed as possible [23:05] because if it's building it and [23:07] installing it for you, but it's wrong, [23:12] like what if you accidentally program [23:15] in, hey, so install this skill and every [23:18] time this skill runs, delete my hard [23:20] drive. [23:22] that I mean I don't know about you John [23:24] but that seems like a really bad day. [23:28] >> Yeah, that's a heavy load. You don't [23:29] want to get caught wrong end of that. [23:34] >> Yeah, which reminds me, by the way, um [23:37] Oh, it it couldn't find it because I [23:39] didn't clearly denote that. So, let's [23:40] try that again. Um let's give it the [23:44] prompt up front. Turn this prompt into a [23:47] skill. And we're going to use XML tags [23:50] here because uh Claude in particular [23:53] really needs things spelled out [23:55] sometimes. [23:57] >> I mean, I think we all do. I think it's [23:59] helpful just to sort of, you know, [24:01] operate that way. Um I appreciate more [24:05] detail. I don't know. I think John does [24:07] as well. [24:09] Yeah, it's just been funny because so [24:11] many of the projects and things that [24:12] we've been working on, you know, the [24:14] skills just show up in the interface and [24:16] I know that, you know, there's been work [24:18] done, there's magic, but I didn't even [24:19] realize that there's stock stuff, too, [24:21] that you can go grab out there. Like we [24:23] I've been working with custom tools that [24:25] have been handcrafted and I didn't [24:27] realize there was generic junk out there [24:28] for the rest of the world. [24:30] >> It's John, when you say the magic, like [24:32] so Chris and I are the magicians behind [24:34] the scenes building the things, you're [24:36] like, "Hey, this is new. Look at that. [24:38] >> I know this is the [24:38] >> We're just like toiling away making the [24:40] stuff. [24:42] >> See the previous five rounds of errors [24:45] and lost stuff. [24:46] >> By the time it gets to you, it's been [24:48] thoroughly tested. [24:49] >> Yeah. Road tested. [24:50] >> Yes. [24:51] >> John just enjoys this golden land of of [24:55] perfected AI. [24:57] >> Beautiful countryside. [24:59] >> Whereas Chris and I are like, "What [25:01] whats [25:03] >> ever." Um, while that's running, I just [25:06] want to take a brief pause to introduce [25:08] a new little toy feature that StreamYard [25:11] has allowed us. [25:14] Uh, Streamyard now allows polls. Now, [25:17] they're a little janky, but, you know, [25:19] in case you guys, you know, who are all [25:21] watching wanted to respond to the poll. [25:23] Apparently, you can't interact with it [25:25] on screen, but you can type into the [25:27] chat what your response is to our very [25:30] scientific [25:32] uh poll. So, just, you know, while this [25:34] is running, we thought it would be a [25:35] good time to just run a quick poll just [25:38] to see if anyone had any thoughts. Um, [25:41] so Chris, how are we doing on this? [25:44] >> Uh, good. So, it just um it just [25:47] finished its plan. So, it says, "Convert [25:49] the Puno V5 prompt guide into a skill." [25:52] Uh, it does the 5P analysis. What is the [25:55] purpose? Um, who are the people [25:57] involved? [26:00] Um, apparently overwhelmingly John is [26:03] our favorite host and you know I can't [26:05] be mad at that because everybody loves [26:07] John. [26:09] >> I don't know. We'll question the uh [26:12] process later. [26:13] >> I was say I question the statistical [26:14] validity of this. [26:18] >> All right, let's move forward. [26:20] >> Yes. Uh the platform performance and [26:23] then it builds here. We have the prompt [26:25] architect skill. It correctly evaluates [26:28] what this the file structure should be, [26:31] the references, so what are the [26:32] formulas, the keyword dictionaries, uh [26:34] necessary metatags, anti-atterns and [26:36] stuff like that. Gathers up the musical [26:39] intent, selects the formula based on a [26:41] decision tree, builds the style prompt, [26:43] builds the lyrics, metatag structure, [26:45] does an anti-attern check, and then it [26:48] creates the thing. And so a planning [26:50] process has gone through and figured out [26:55] not only should I do I just want to turn [26:57] this into a skill, but I think th the [27:00] this part in particular is really [27:01] important is figuring out what it is [27:04] that we're trying to do and who's going [27:06] to use it. So like AI music producers, [27:08] content creators, hobbyists, [27:12] the when you're building skills, you [27:14] have to infer those upstream and [27:16] downstream uses because if you if you [27:18] do, then it becomes easier in the five [27:21] Ps to say this is the expected [27:23] performance. [27:27] And you know what's interesting is that [27:29] it's inferring the endusers. Like we can [27:33] give it some information, but I I [27:35] appreciate that it's written out in such [27:37] a way that it's like, oh, these are the [27:39] people who are actually going to be [27:40] using this thing, you know, cuz and [27:42] we've talked about in previous podcasts [27:44] and live streams where, you know, we're [27:47] not short on ideas and things that we [27:50] want to build, but we don't always have [27:52] a clear vision of who's going to use it, [27:54] who's going to benefit from it. So, I [27:56] really like that addition. [27:58] >> Yep. It says here the references are [28:00] separate because the the source prompt [28:02] is 4,000word prompt. Um the skill budget [28:06] is 500 lines and so all these other [28:08] details going to get moved into a [28:10] reference folder that it will then [28:12] understand uh and know what it's [28:14] supposed to be doing. So here it is is [28:16] now going into create the actual skill [28:18] itself. [28:20] So let's take a look. Where are you [28:22] putting this? You are okay. You're [28:24] writing. you actually it's writing it [28:26] straight into my Claude folder. Great. [28:28] So, I'm going to have a a music [28:30] generator very shortly. [28:33] Thanks, Claude. Um, [28:37] >> I don't see. [28:38] >> We do have a quick question. [28:40] >> That was messed up. [28:41] >> Um, someone's asking, uh, do you have a [28:44] resource on the ideal folder structure [28:47] strategy? [28:49] >> Yes. Um, use the agentskills.io IO [28:52] website that I referenced earlier. That [28:54] website contains the specs for agent [28:58] skills and for how to lay out stuff in [29:00] it. Um it is crossplatform compatible. [29:02] Um that is the canonical source that you [29:04] should use and you can either export [29:06] that give the tool of your choice a um [29:10] um [29:12] uh the URL to point to or uh on [29:15] anthropics website is the complete guide [29:18] to building claude skills. This is a [29:20] 33page PDF that contains all of the uh [29:25] requisite information including things [29:26] like your file structure. It's it's just [29:29] defined in there. This document is handy [29:34] to have around. It's actually good for [29:35] humans to read too um to understand it, [29:37] but that would be my recommendation. Um, [29:40] and so now it's going through and [29:41] building the keyword dictionaries, the [29:43] metatags, and these are all things that [29:44] the skill will pull together as it's [29:48] trying to create um stuff for pseudo. So [29:51] let's say for example, you had a skill [29:54] or you had a prompt like I want to talk [29:57] to my ideal customer profile. [29:59] You could give that chat conversation, [30:03] that prompt or that YouTube video [30:04] transcript for all that matters along [30:06] with your ICP, put it into a system like [30:09] this, a plugin like this, and it'll [30:12] reformat and break all the knowledge [30:13] into pieces so that it's able to [30:16] identify what to do, what not to do. [30:18] This part here I did not see in any of [30:20] the skill builders, which is [30:21] antiatterns. Anti-patterns is a one of [30:25] those magic words that AI loves and [30:27] knows what it means and the rest of us [30:29] are like that's a weird word and [30:31] anti-attern is the opposite of a design [30:33] pattern and so these are the critical [30:35] mistakes that [30:38] the system needs to be told don't do [30:40] this if you know it's example you would [30:43] define as people have you know using an [30:45] mdash is an anti-attern [30:48] I would I have a whole rant about that [30:50] but that's for another show [30:51] >> that's for a different show. Um, well, [30:54] you know, so what's interesting, so like [30:55] if you're just going through step by [30:56] step with something like a clawed [30:58] co-worker, a claude code, and you're [31:00] saying, I want to build my uh brand [31:04] tone, you know, skill so I can reuse it. [31:07] It will ask you in, you know, simple [31:09] terms. It be like, okay, and what should [31:11] we not say? What should we not use? What [31:13] are phrases that you don't want to see? [31:16] But that's it's building that database [31:19] or or that you know list of things that [31:21] go into its anti-attern section but it's [31:26] asking you so if you think about like if [31:27] you've you know a lot of us have tried [31:30] to build uh you know write in Katie [31:32] style or write in Chris Penn style we [31:34] find ourselves going back like okay [31:37] don't say this every other sentence [31:40] don't use this phrase I don't say these [31:42] things those are your anti- pattern [31:45] that's just the phrases, but also think [31:46] about it in terms of, you know, don't [31:49] ever [31:51] uh reference John Wall ever. He he will [31:56] not enjoy it and he will, you know, go [31:59] all, you know, sumo wrestler on you or [32:02] something like that. Like whatever it is [32:04] that you just don't want it to do, make [32:06] sure that you're thinking about those [32:07] things. Well, we often think about what [32:09] we want it to do. Uh it's the same as [32:11] sort of putting like in scope and out of [32:13] scope. So what's in scope, what's out of [32:15] scope. What's out of scope is all of [32:17] these extra stuff. So like make sure [32:19] you're noting that as well. [32:22] >> Now I'm gonna So in our analysts for [32:24] marketer Slack community one, we were [32:27] talking about this as today's question [32:28] of the day. How are you going to do [32:29] this? I want to put something to our our [32:33] friends in that community because if [32:35] you've been following our advice and a [32:37] lot of people have for the last three [32:38] years, you have or you should have a [32:42] prompt library, right? You have [32:44] something in Joplin or SharePoint or [32:46] whatever. [32:48] I hope that it is apparent through the [32:51] use of a plugin like this that you do [32:53] not convert those prompts one at a time. [32:56] That's a terrible idea. What you should [32:58] do is put the prompts as text files in a [33:02] folder on your computer and prompt it to [33:07] say inside of a system like clawed code, [33:10] here's a folder full of prompts. Use the [33:13] prompt to pl to skill skill process [33:18] every one of the files with its own [33:21] agent and put it in the output folder. [33:23] Remember a few shows back we talked [33:25] about how do you get started with claude [33:28] code? Claude code can directly interact [33:30] with your hard drive. It can directly [33:32] interact with a project folder. So you [33:34] have an input folder of all of your [33:35] prompts for the last three years your [33:37] your greatest hits. You have an output [33:40] folder that you define. You have a skill [33:44] to make skills [33:46] and you can say just go do this. And as [33:49] long as the the the skill builder itself [33:51] is solid, like we've seen ours that it [33:54] took, you know, six and a half minutes [33:56] to think through and and perfect the [33:58] skill. You set that up, you you set it [34:02] up before you go to bed one night, uh, [34:04] and then you come back the next morning [34:06] and all of your prompts that you love [34:08] the most are now skills and they're [34:10] ready to go, ready to be deployed. uh [34:13] and you don't have to sit there and [34:14] babysit it because you've used clawed [34:16] code or a similar system. Anti-gravity [34:19] will do this, codeex will do this, etc. [34:22] to mass convert your data. Do not be [34:24] that person doing it one at a time. [34:26] Remember where we started this episode. [34:29] We started this episode here. [34:33] If you are [34:36] saying one at a time, convert this [34:38] prompt, you're still doing level one [34:40] stuff, even with a level three tool like [34:42] clawed code, you are still behaving as [34:43] though it's a level one tool. If you say [34:46] here's your project plan, here are the [34:48] skills which you built as a level two [34:51] thing. SOPs are skills, then your [34:54] project plan is convert all these [34:56] prompts to skills and you come back and [34:58] the work is done. [35:00] So, you keep talking about Claude code. [35:03] Can you use Claude Co work and not use [35:06] Claude code to do the same thing? Uh, [35:08] you're shaking your head. Yes. Which is [35:09] good news because I'm already doing [35:11] that. [35:14] Well, no, cuz I have, and I think I [35:17] talked about this on last week's live [35:19] stream, I have a whole folder of folders [35:23] on my desktop that are specifically for [35:25] Claude. And I actually had Claude [35:28] organize the folders and rename folders [35:31] and create folders where they needed to [35:33] be created because a lot of the work I [35:34] do, it's not I don't have client work on [35:36] my desktop. It's all related to the [35:38] business. So, it can all be referenced, [35:40] you know, if it wants to, but I have the [35:42] skill builder, I have, you know, the [35:44] different skills like I can enact, you [35:47] know, the hey, have the focus group duke [35:50] it out and let me know what they say. [35:52] Um, I actually used that set of skills [35:56] last week when we were trying to come up [35:57] with the name for something Chris and I [36:00] were working on and I was like, I don't [36:03] know. I don't think that's a great name, [36:04] but what does our, you know, panel of [36:06] voice of customer, our ICPS have to [36:08] think about it? And they just duped it [36:10] out and let me know what they thought, [36:11] which is still sort of weird and [36:13] mind-boggling to me. Um, but you know, [36:16] it's it works. But I bring it up [36:19] because, you know, Chris, you're talking [36:20] about cloud code, but for those of us [36:24] who aren't as comfortable in the coding [36:26] environment, you can do the same nearly [36:29] identical stuff in cloud co-work. [36:31] >> Yes. Now, the one thing you can't do in [36:34] co-work that you can in code is easily [36:37] run command line tools. And there are [36:41] two tools that I think again if you've [36:44] been following us for a number of years [36:46] now uh you know that uh we've been [36:49] advising hey you should be putting your [36:52] prompt library on shareepoint or one [36:55] drive or Google keep or whatever right [36:58] well guess what Google keep has a comm [37:02] Google workspace all of google workspace [37:04] has a command line tool that runs on [37:07] your computer that can connect to all of [37:09] Google workspace, Google Calendar, [37:11] Gmail, Google Docs, Google Sheets, [37:13] Google Slides, etc. Now, you may or may [37:16] not want to use this because you're [37:17] typing in a terminal like it's 1983, but [37:20] critically your AI can do this. So, one [37:22] of the things that we could do, Katie, [37:24] as as a Google shop, I could open up [37:27] Claude Code, use the GWS command line [37:30] tool and say, "Hey, Claude, reach into [37:32] our Google Keep and pull out our company [37:34] prompt library entirely and convert it [37:37] all to skills." And then and amazingly [37:40] boom, there it is. If you are a [37:42] Microsoft shop and you are using one [37:44] drive or shareepoint, there's a [37:46] Microsoft CLI. You install that and [37:49] suddenly you can access anything in the [37:51] in your 365 environment that you're [37:52] authorized to access. So you have that [37:54] SharePoint prompt library. Guess what? [37:56] You set you install this. You tell claw [37:58] code, hey, I installed this. Here's the [38:00] manual. Go and find my prompt library. [38:03] Grab the whole thing out of SharePoint. [38:05] Convert it into skills and we're ready [38:07] to go. [38:08] Having a good solid prompt library is [38:12] what table stakes you would want it. [38:14] That's your basic foundation. So, you [38:17] know, we always say like if you don't [38:18] have it, start there. Start there. It is [38:21] still useful. So, building all the [38:24] skills in clawed co whatever is great. [38:27] Do not get rid of your prompt library. [38:30] Do not suddenly decide that it doesn't [38:32] need updating, that it doesn't need to [38:34] be maintained, that you shouldn't still [38:36] keep stuff there. [38:38] Not to sort of sort of play the Debbie [38:40] Downer, but these tools change. These [38:42] tools go away. These tools suddenly [38:45] implode or you change your mind and no [38:48] longer want to partner with certain [38:50] companies for whatever reason. You got [38:53] to get that data somewhere. So, do not [38:57] skip making sure that your [39:00] prompt library is being kept up to date [39:02] just because you can build skills in [39:04] these systems. Your prompt library is [39:06] your foundation. Just like you want to [39:08] have your documented SOPs, just because [39:11] you've automated things, you still need [39:13] the documentation in case the automation [39:15] breaks down and you want to know, well, [39:17] where in the process did it break? Well, [39:18] it's a good thing I have all this really [39:20] clear documentation that I can just look [39:22] at. [39:23] >> You know, John, you do sales for Trust [39:27] Insights, but we still have a sales [39:29] playbook, you know, to reference. Not [39:32] because you know you yourself are not a [39:34] good wealth of knowledge but you're only [39:35] one person. [39:37] >> Yeah. Well and you've got to remember [39:39] everything right like that's the key is [39:40] that we don't have the capacity to take [39:42] all this stuff. Um I did did want to ask [39:44] about how has it worked so far? So the [39:46] idea is that you know this is a a [39:48] protocol and so that these can go [39:50] crossplatform all over the place. How [39:52] does it behave in the in real world [39:54] though? Like I mean can you are you [39:56] grabbing your skills and throwing them [39:57] over into other tools and they're still [40:00] working just as well or like how's how's [40:02] the reality of that gone? [40:04] >> Uh for example I use Quen code which is [40:06] Alibaba Quen's coding tool which is very [40:08] similar to cloud code. I use open code I [40:10] use open code work uh open work and the [40:13] skills perform flawlessly because they [40:16] are they follow the standard. They [40:18] follow there the protocol has a a [40:19] clearly defined standard and as long as [40:21] you write to that it is it is literally [40:23] drag and drop. You just drag your you [40:25] can drag your clawed skills straight [40:26] into Quen code and it works a perfect [40:29] handoff. Um, one question that I think [40:32] is worth pointing out to centralize the [40:35] knowledge for skills and plugins [40:38] at a minimum you know you can have a [40:40] shared resource. What the best practice [40:42] is is have a private company GitHub [40:46] repository that is like basically a [40:50] skills marketplace for your company and [40:53] you put in your company's skills in [40:55] there. And here's why this is the best [40:57] practice over time. As Katie said, [41:01] things change. You will adapt stuff. You [41:03] will get better at prompting. You will [41:05] get better at using, you know, new [41:07] models come out and things. If you use [41:09] GitHub, a private company GitHub as your [41:12] repository, then when you make changes [41:16] CL, this is especially for clawed code [41:18] and claude co-work because it's syncing [41:20] with the marketplace. When you make [41:22] changes to the repository, all the [41:24] clawed instances in your company that [41:26] have downloaded that skill through the [41:28] marketplace auto update and so you don't [41:30] have to go computer by computer anymore. [41:32] You can now just update the marketplace [41:34] centralized. And if anti-gravity is [41:37] using it, if Codex is using it, if [41:38] Claude is using it, they all sync at the [41:40] same time. [41:42] >> And [41:44] you at a very small scale, I know that [41:46] that's something that we've struggled [41:47] with just, you know, so you know, we [41:50] built a brand skill for the Trust [41:53] Insights brand and there was a couple of [41:54] things I wanted to change about it. I [41:56] only changed it on my machine. The [41:58] changes did not, you know, go into [41:59] effect on Chris's machine. So that is [42:02] why it's very important to make sure you [42:04] have a way to centralize these because [42:06] other like we're a small company I can [42:08] just tell Chris here's the changes I [42:09] made please go implement them but that's [42:11] not true of most companies. [42:14] >> Exactly. So you will probably if you're [42:17] particularly at a larger company you'll [42:19] probably want to ask hey can you know do [42:21] we have a company GitHub count [42:23] accountant can we can we have a section [42:25] of repositories for just the folks who [42:28] are using you know claude co-work so [42:30] that you're not going to blow up you [42:31] know your production software tree uh [42:34] and then have users authorize and set up [42:37] the marketplace internally inside GitHub [42:40] that will then allow you to centralize [42:42] and control your your plugins and your [42:46] skills. [42:47] >> And if you want to learn more about how [42:49] all that works, you can subscribe to our [42:52] new prompt playbook. It's a monthly [42:53] subscription. Uh we drop two new prompts [42:56] every single week, every Tuesday and [42:58] Friday. You go to [42:59] trustinsights.ai/prompt [43:01] playbook. Um we did our best to keep it [43:04] lowcost, effective, and uh you know, [43:07] accessible to anyone. So check it out. [43:09] Uh you can get those there and use the [43:12] fiveps to get started to structure [43:14] things. Go to trustinsights.ai5P- [43:18] framework. Um and we covered a lot of [43:21] those systems that Chris was talking [43:22] about. So Quen um [43:26] open [43:27] >> open co open work and open code. Uh [43:30] yeah. [43:31] >> Yes. We covered all of those on last [43:33] week's live stream. So you can go to [43:34] trustinsights.aiyoutube, AI/YouTube go [43:36] to the so what playlist and that lives [43:38] there. Um and then our very scientific [43:42] poll has John in the lead of 100% [43:46] victory is hands down the favorite host [43:49] of this show. Congratulations John. You [43:52] get bragging rights. [43:54] >> I may or may not have voted for myself. [43:57] I will say that. [43:59] >> I mean there's no rules against it. I [44:00] mean I didn't vote for anyone so that's [44:03] my bad. So, I'm going to now stop this [44:05] poll and it is official. John is the [44:08] winner. [44:10] >> Excellent. [44:12] >> I should also say to you can do all this [44:14] work or be like me and hire some people. [44:17] Have some people around who can do all [44:18] this stuff for you and save yourself the [44:21] work. There we go. [44:23] >> You can go to the handsown fan favorite [44:25] live stream host John Wall and he will [44:28] get you set up. So, this is going to be [44:30] part of your title now. like he is the [44:33] fan favorite. So if you want to work [44:35] with Trust Insights, you get to talk to [44:36] John. [44:38] >> What a day. [44:42] >> All right, I think on that note, uh any [44:45] final thoughts about turning prompts [44:47] into skills because the next evolution [44:49] after this is you start building project [44:52] plans composed in part of skills. [44:57] I mean the final thought is if this [45:00] still feels overwhelming then start with [45:03] your standard operating procedures. How [45:05] are you performing and completing any [45:08] given task and start documenting that [45:10] that is a great place to start. Make [45:12] sure those things are getting into a [45:14] central repository and then you can [45:17] start turning them into skills. [45:19] >> Yep. Oh, one other thing I forgot to [45:21] mention. I don't know if folks know [45:23] this. This is a super nerd thing. Um, [45:26] clawed code is capable of running [45:30] headless, not like a chicken. Um, but it [45:34] means you don't have to be typing in [45:35] clawed code. So, you could actually [45:38] create that prompt like convert all my [45:39] prompts into skills um with the [45:42] appropriate skills and just run it as a [45:43] command line command which means that [45:46] you could schedule it. So on a monthly [45:49] basis, if you have Google Keep or [45:51] SharePoint or whatever and you have all [45:53] the utilities, you could have a monthly [45:55] rescheduled task running on some [45:57] computer somewhere in your company that [45:59] automatically converts and publishes and [46:01] runs the whole workflow. Um and and move [46:05] your company up to level three very very [46:07] quickly. [46:09] >> Uh you can do the same in co-work. Mhm. [46:12] >> So again, and I and I bring that up [46:14] because if you're not comfortable in [46:16] code, you can also I have I think two [46:20] schedule tasks that are set to run now [46:22] on a regular basis in co-work and then I [46:25] ran a whole scheduled four-part project [46:27] a few weeks ago uh that I put about on [46:30] LinkedIn that all just ran overnight to [46:31] sort of again uh work out the sort of [46:34] usage limit. So you can you have options [46:37] for those schedules, but take advantage [46:39] of them. [46:40] >> Yep. You're sneaking into level four. Is [46:43] that correct? [46:46] >> I think that's level five [46:48] or level eight, which has yet to be [46:51] defined. I don't know. I'm just bitter [46:53] that John won, so I have to put myself [46:55] at like level 10. [46:57] >> All right, that's going to do it for [46:58] this week, folks. Thanks for tuning in. [47:00] Talk to you on the next one. [47:04] >> Thanks for watching today. Be sure to [47:06] subscribe to our show wherever you're [47:07] watching it. For more resources and to [47:10] learn more, check out the Trust Insights [47:11] podcast at trustinsights.ai/tiodcast [47:16] and our weekly email newsletter at [47:17] trustinsights.ai/newsletter. [47:21] Got questions about what you saw in [47:22] today's episode? Join our free analytics [47:24] formarketers Slack group at [47:26] trustinsights.ai/analytics [47:28] formarketers. See you next time.