The 7 ChatGPT Work Skills I Use Every Day  🔗 **Source:** [The 7 ChatGPT Work Skills I Use Every Day](https://www.youtube.com/watch?v=Dkdgut6i2xY) > [!note]+ Document Notes > tags: [ai-skills, chatgpt, workflow-automation, prompt-engineering, ai-agents, productivity-tooling] > author: Paul J Lipsky > source: YouTube > reference: https://www.youtube.com/watch?v=Dkdgut6i2xY > date created: 2026-08-31 10:00 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: The 7 ChatGPT Work Skills I Use Every Day (Copy Them) > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [The video under analysis](https://www.youtube.com/watch?v=Dkdgut6i2xY) — Paul J Lipsky, 17:20, published 2026-08-31 - [Pushover](https://pushover.net/) — the notification service the LMK skill depends on - [Matt Pocock](https://www.mattpocock.com/) — original author of the Grill Me skill, credited in the video - [Remotion](https://www.remotion.dev/) — the plugin whose 12 bundled skills close the video ## 📝 Summary - **What happened.** Paul J Lipsky walks through seven ChatGPT Work "skills" — reusable saved instruction sets invoked with `/name` — that he says he uses almost daily, plus a bonus. In order: **visualize** (preloaded; renders responses as interactive artifacts), **LMK** (push notification on task completion via Pushover), **handoff** (packages project context for a different AI), **ask** (drives another AI through browser-use), **grill me** (forces an interview before the model acts), a **multi-step skill** chaining Vimeo → Gmail draft → Obsidian update → notify, and **Remotion's** 12 bundled plugin skills. Bonus: **hatch**, which generates desktop "pets." He offers all prompts free with no email gate. - **Why it matters.** This is a practitioner demonstrating that the useful unit of AI work has shifted from the prompt to the *saved, composable, invocable* skill — and that skills compose with browser control, local file access, plugin integrations, and scheduled automations into workflows that run unattended. Anyone already authoring skills will recognize the pattern; the specific compositions are the value. - **Key stakeholders.** ChatGPT Work users on a tier with plugins, skills, browser use, and local file access. Skill authors. Matt Pocock, credited as the origin of Grill Me. Pushover and design.com, the two third-party services with a commercial relationship to the content. Remotion, whose bundled-skills model the video uses as its closing argument. - **Risks and tradeoffs.** Substantial and entirely unaddressed. The LMK setup instructs viewers to paste a Pushover **API token into a chat context**. The `ask` skill has ChatGPT drive Grok and Gemini through a browser using the viewer's own logged-in subscriptions — a terms-of-service question the video does not raise. The multi-step skill combines browser control, Gmail write access, and local filesystem writes in one unattended chain. None of this is framed as carrying risk. - **Bottom line.** Genuinely useful, well-paced, concretely demonstrated, and honestly sourced where it borrows. The `handoff` and `grill me` patterns are the two most transferable ideas and both are worth stealing. The video's blind spot is that every skill it demonstrates expands what an autonomous system can do with your credentials, your inbox, and your filesystem, and it treats that expansion as pure upside. Watch it for the compositions; supply your own threat model. ## 📖 Main Content **The core claim, and what is actually true about it.** Lipsky's framing is that skills convert one-off prompting into repeatable capability: *"you have to reexplain your workflows to chat GPT every time you use it. And that costs you time and gives you inconsistent results."* That is an accurate description of a real problem, and the proposed fix — save the instructions, invoke them by name — is the right shape. Where the framing overreaches is the Matrix analogy: *"with one simple upload, Neo was suddenly a kung fu master."* A skill does not add capability the model lacks. It removes the cost of re-specifying capability the model already has, and it makes the specification consistent across invocations. Those are worthwhile and they are not the same thing. The distinction matters practically, because it predicts which tasks skills will help with (repeated, well-specified, currently re-explained) and which they will not (anything the model cannot do at all). **Mechanics, briefly.** Skills live under Plugins → Skills in the ChatGPT desktop app; some ship preinstalled, others are user-created. The recommended authoring path is conversational rather than form-based — open a new chat, describe the skill you want, and the model writes and saves it. Invocation is `/name`. This is the same authoring ergonomics that make skills stick: the cost of creating one has to be lower than the cost of re-explaining the workflow, or nobody creates them. **The seven, and which ones earn their place.** *Visualize* is preloaded and, by Lipsky's account, underused. He points it at a local folder of family budget files and gets an interactive breakdown — core spending, savings, planned cash out, a monthly/annual toggle, an itemized table — where the unskilled response was prose. This is the clearest before/after in the video and the least risky item in it. *LMK* solves a real irritation — ChatGPT not reliably notifying you when a long task finishes — by routing a completion signal to Pushover, a $5 one-time notification service. It is also the item that should give a careful viewer pause, and I will return to it. *Handoff* is the strongest idea in the video and the one most worth generalizing. It packages the state of an in-progress project — *"the goal of the project, what's already been done, the decisions you've made, any files or links that it needs, and exactly what should happen next"* — into a portable brief for a different model. His worked example is honest about hitting a wall: a phone app project stalled in ChatGPT, handed to Claude, unblocked, handed back. That is a working description of model-switching as a deliberate technique rather than an admission of failure, and the five-part structure is close to a general-purpose spec for context transfer between any two agents. *Ask* is handoff's lightweight sibling: rather than moving a whole project, have ChatGPT open a browser, visit another AI, and ask one question. His uses are Grok for real-time X sentiment (ChatGPT lacks live X access) and Gemini for image generation he prefers. He is explicit about the motivation — *"it takes advantage of the subscriptions I'm already paying for"* — and about why he prefers it to doing it by hand: only the automated form composes into scheduled tasks. *Grill Me* forces the model to interview you before acting, surfacing *"the assumptions it's making, the different decisions that you want, and the missing details you haven't thought through yet,"* and blocks execution until both parties agree the spec is settled. Lipsky credits Matt Pocock as the original author, says Pocock's version is *"quite intense,"* and offers a scaled-down variant. The attribution is unprompted and specific, which is worth noting in a genre where borrowed prompts usually arrive unsourced. *The multi-step skill* is the most operationally ambitious. One invocation plus an Obsidian link triggers: open Vimeo in a browser, locate the video, copy the link, draft an email to the sponsor, update the status and schedule in Obsidian, then notify him to review and send. He is appropriately clear that the specific skill is not reusable by viewers and that the *pattern* is the point — find a rote multi-step workflow and encode it. *Remotion's bundled skills* close the video with a structural observation rather than a recipe: installing the plugin installs twelve skills, and the user never needs to know their names because the model selects among them. That is a real shift in how capability gets distributed — the skill becomes a packaging format, not just a personal artifact. **What the video does not say, and should.** Three risks are demonstrated on screen and named nowhere. First, **credential handling**. The LMK walkthrough has the viewer paste a Pushover user key and API token into a chat. Whatever the provider's retention and training policy, a long-lived API token now exists in a conversation transcript. There is no discussion of scoping, rotation, or what happens if that context is later shared, exported, or used to train. Second, **terms of service**. Having ChatGPT drive a browser to operate Grok and Gemini through the viewer's own authenticated sessions is exactly the automated-UI-access pattern that AI providers' terms commonly restrict. Lipsky's framing — reuse the subscriptions you already pay for — is economically sensible and legally unexamined. A single sentence acknowledging the question would cost nothing. Third, **compound blast radius**. The sponsor-review skill holds, simultaneously, browser control, Gmail write access, and local filesystem write access to an Obsidian vault, running unattended on a Mac Mini on a schedule. Each capability is defensible alone. Together, and automated, they describe a system that can send mail as him and modify his notes without a human in the loop. The video presents this as the payoff, which it is, without noting that it is also the risk surface. **Contested and uncertain.** *"you'll never have to repeat yourself ever again"* is an absolute that skill drift, model updates, and changing workflows will falsify. The claim that skills produce *"better, more consistent outputs every time"* is asserted without a comparison beyond the one visualize demo. And the video's central premise — that these are the seven he uses *every day* — is unverifiable, though the specificity of the sponsor-review example argues for genuine use rather than manufactured content. ### 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=0) Hook and the skills premise - Cold open contrasts an unskilled summary against a skilled one — show, don't tell, and effective. - Core value claim: without skills you re-explain workflows every time, costing time and consistency. - The Matrix/Neo analogy overclaims: skills reduce specification cost, they do not add capability. - Free prompts promised, no email gate. Explicit thumbs-up and watch-to-the-end ask. #### ⏱️ [00:52](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=52) Where skills live and how to make one - Desktop app → Plugins → Skills. Some preinstalled, some personal. - Two creation paths: the Add button, or — recommended — describe the skill in a new chat and let the model author and save it. - Conversational authoring is the important detail: creation cost must undercut re-explanation cost or the habit never forms. #### ⏱️ [01:47](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=107) Skill 1 — visualize (preloaded) - Invoked with `/visualize`. Turns prose responses into interactive artifacts. - Demo: local folder of family budgeting files → interactive breakdown with monthly/annual toggle and itemized cost table. - Clearest before/after in the video; lowest-risk item. #### ⏱️ [03:36](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=216) Skill 2 — LMK (let me know) - Solves unreliable mobile completion notifications by routing through **Pushover** (free trial, $5 one-time). - Setup: obtain user key, create an application, obtain an API token, **give both to ChatGPT**. - ⚠️ A long-lived API token is pasted into a chat context. No discussion of scope, rotation, retention, or export risk. #### ⏱️ [05:21](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=321) Composing skills into scheduled automations - Running automation on a Mac Mini monitors trending X topics and pushes a notification only when something warrants attention. - The general pattern: skill + schedule + notification = an attention filter that runs without you. #### ⏱️ [06:16](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=376) Sponsor — design.com - Verbally disclosed as sponsorship at the segment's start and end. Adequate disclosure. - Roughly one minute of a seventeen-minute video; not disproportionate. #### ⏱️ [07:11](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=431) Skill 3 — handoff - Packages: goal · what's done · decisions made · files and links needed · what happens next. - Worked example: stalled iPhone app project → handoff → Claude unblocks it → return handoff → back to ChatGPT. - Best idea in the video. The five-part structure generalizes to any agent-to-agent context transfer. #### ⏱️ [08:57](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=537) Skill 4 — ask (browser use) - Lighter than handoff: open a browser, visit another AI, ask one question, bring the answer back. - Uses: Grok for live X sentiment (ChatGPT lacks real-time X access); Gemini for preferred image generation. - Rationale: reuse subscriptions already paid for, and — the real reason — only the automated form composes into scheduled tasks. - ⚠️ Automated UI access to third-party AI services via authenticated sessions. ToS question unraised. #### ⏱️ [10:48](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=648) Skill 5 — grill me - Forces an interview before action; surfaces assumptions, decisions, and missing details; blocks execution until both agree. - **Credits Matt Pocock** as the original author; describes that version as "quite intense" and offers a lighter fork. - Unprompted, specific attribution — notable in a genre where borrowed prompts usually travel unsourced. #### ⏱️ [11:45](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=705) Skill 6 — multi-step (sponsor video review) - One invocation + an Obsidian link → Vimeo lookup → grab link → draft sponsor email → update Obsidian status and schedule → notify for review. - Composes browser use + Gmail plugin + local file access in a single unattended chain. - He explicitly says the pattern, not this skill, is the transferable part. Correct framing. - ⚠️ Compound blast radius: can send mail as him and modify his vault without a human in the loop. #### ⏱️ [14:37](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=877) Bonus — pets and hatch - `/pet` spawns a persistent desktop companion that survives app switching and signals task completion. - `/hatch` generates custom pets from description or reference images. His is a paperclip named PIP. - Genuinely a completion-notification affordance wearing a costume; the ambient-presence idea is more interesting than the novelty. #### ⏱️ [15:29](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=929) Skill 7 — Remotion's bundled skills - Installing the Remotion plugin installs **12 skills** (best practices, captions, docs, maps, others). - Key point: the user never invokes them by name; the model selects among them automatically. - Skills as a *packaging format* for plugin capability, not just a personal artifact. The most structurally interesting observation in the video. #### ⏱️ [16:26](https://www.youtube.com/watch?v=Dkdgut6i2xY?t=986) Outro - Pitches a companion video on a skill that edits his YouTube videos. - Standard end-screen funnel; no new content. #### Transcript quality note - Auto-captions render "ChatGPT" variously as "chat GBT," "chatbt," "Chad GPT," and "chat GPD"; "Matt Pocock" appears as "Matt PCO"; "Grok" as "Grock." All corrected here. Worth knowing if searching the transcript later. ## 📚 References - [Paul J Lipsky, "The 7 ChatGPT Work Skills I Use Every Day (Copy Them)," 2026-08-31](https://www.youtube.com/watch?v=Dkdgut6i2xY) — the source under analysis - [Pushover](https://pushover.net/) — notification service underpinning the LMK skill; free trial, $5 one-time - [Matt Pocock](https://www.mattpocock.com/) — credited originator of the Grill Me pattern - [Remotion](https://www.remotion.dev/) — programmatic video plugin shipping 12 bundled skills - [design.com](https://design.com/) — disclosed sponsor of the video's mid-roll segment ## 🗣️ Paraphrase - **Opening.** A default ChatGPT Work summary looks one way; the same request with a skill applied looks considerably better. The right skills produce better and more consistent outputs every time. The video covers seven he uses almost daily, supporting interactive visualizations, more effective research, and motion graphics for his videos, and the prompts are free. - **Why skills.** Skills unlock capabilities already inside ChatGPT — likened to Neo's instant kung fu upload in The Matrix. Without them you re-explain workflows every session, which costs time and yields inconsistent results. Because skills are repeatable instructions specifying exactly how a task should be done, the repetition ends. - **Getting started.** In the desktop app, Plugins → Skills lists what is installed, some preloaded and some personal. You can add via the Add button, but the better route is to open a new chat in ChatGPT Work, describe the skill you want, and let it be created and saved for reuse. - **Visualize.** A preloaded and easily missed skill, invoked by typing `/` and its name, that converts responses into visual output. Pointed at a folder of family budgeting and spending files, it produced an interactive view of core spending, savings, and planned cash out, with a monthly/annual toggle and a table of known costs — far easier to read than the prose version. - **LMK.** Addresses never receiving the phone notification when a task finishes. Configured with Pushover — download, log in, collect a user key, create an application, collect an API token, and give both to ChatGPT. A linked page carries the setup prompts with no email required, in exchange for a thumbs up and watching to the end. Invoked with `/lmk`, and used inside automations such as a Mac Mini job that watches trending X topics and pushes only what warrants attention. - **Sponsor.** design.com, an all-in-one AI design platform for logos, business cards, social graphics, and websites, which propagates a generated logo's colors and fonts across subsequent assets, and can build around an uploaded existing logo. - **Handoff.** For people using several AI tools, a handoff packages what matters — goal, progress, decisions, needed files and links, and next steps — into a compact brief for another model. A stalled iPhone app project was handed to Claude, which resolved the blocker and produced a return handoff back to ChatGPT. - **Ask.** Where handoff moves a project, ask poses a single question, instructing ChatGPT to open its browser and consult another AI. Used to query Grok for X sentiment ChatGPT cannot see, and Gemini for preferred image generation. The advantages are reusing existing subscriptions and, unlike doing it manually, composing into automations with ChatGPT as central delegator. - **Grill me.** Prevents ChatGPT from charging ahead on a partially understood request by forcing it to interview you first, exposing assumptions, decisions, and unconsidered details, and withholding work until both sides agree. Originally by Matt Pocock; his version is intense, so a lighter variant is offered. - **Multi-step.** Skills can chain sequences. His sponsor video review skill, given an Obsidian link, visits Vimeo, retrieves the video link, drafts the sponsor email, updates the Obsidian schedule and status, and notifies him to review and send — combining browser use, the Gmail plugin, and local file access. Viewers should look for their own rote multi-step workflows to encode. - **Pets.** `/pet` produces a desktop companion that persists across apps and signals when a task needs attention or finishes; `/hatch` creates custom ones from description or reference images. - **Remotion.** Installing the Remotion plugin brings 12 skills covering best practices, captions, docs, and maps. Users need not know them by name — ChatGPT selects and installs what it needs automatically, a pattern common to many plugins. - **Close.** A companion video covers a skill that edits his YouTube videos. ## 📰 Newsletter **The unit of AI work isn't the prompt anymore. It's the saved skill.** Paul Lipsky's seventeen minutes on ChatGPT Work skills is worth your time, and not for the reason the title suggests. The list of seven is fine. The two ideas underneath it are better than the list. **Idea one: `handoff`.** When a project stalls in one model, don't start over in another — have the first one write a brief. Lipsky's structure is five parts: *the goal, what's already been done, the decisions you've made, any files or links needed, and exactly what should happen next.* He hit a wall building an iPhone app in ChatGPT, handed off to Claude, got unstuck, and Claude wrote a handoff back. That is model-switching as technique rather than as defeat, and the five-part shape generalizes to any two agents you want talking to each other. If you take one thing, take this. **Idea two: `grill me`.** Force the model to interview you before it builds anything — surface its assumptions, make you state the decisions, find the details you skipped — and block execution until you both agree the spec is settled. Credit where he gives it: this is **Matt Pocock's** pattern, and Lipsky says so unprompted, then admits Pocock's version is "quite intense" and offers a lighter fork. Unsourced prompt-borrowing is the norm in this genre; this isn't that. The rest is solid and varied. `visualize` ships preloaded and most people never touch it — he points it at a folder of family budget files and gets an interactive monthly/annual breakdown where prose used to be. `LMK` pushes a phone notification when a long task finishes. A multi-step skill takes one command and an Obsidian link, then hits Vimeo, drafts the sponsor email, updates his vault, and taps him to review. And the closer is the structurally interesting one: install the Remotion plugin and you get **twelve skills** you never invoke by name, because the model picks among them. Skills as a packaging format for capability, not just a personal shortcut. **Now the part the video leaves out, and it is not small.** The LMK walkthrough has you paste a **Pushover API token into a chat**. Not a password manager, not an environment variable — a conversation transcript. Nobody mentions scoping, rotation, or what happens when that context gets exported or used for training. The `ask` skill has ChatGPT drive a browser into Grok and Gemini using *your* logged-in sessions. His words: "it takes advantage of the subscriptions I'm already paying for." Economically sensible; also precisely the automated-UI-access pattern that AI providers' terms tend to restrict. One sentence acknowledging that would have cost nothing. And the sponsor-review skill holds browser control, Gmail **write** access, and local filesystem writes to an Obsidian vault, running unattended on a schedule. Each is fine alone. Chained and automated, that is a system that can send mail as you and edit your notes with nobody watching. The video presents this as the payoff — which it is — without noting it is also the blast radius. **The framing to resist:** "with one simple upload, Neo was suddenly a kung fu master." A skill doesn't add capability the model lacks. It removes the cost of re-specifying capability the model already has, and makes that specification consistent. Useful! Not the same thing. The distinction tells you what skills will actually help with — repeated, well-specified work you keep re-explaining — and what they won't touch. **Worth noting for the WordPress crowd:** the pattern of a plugin shipping skills that the model auto-selects is going to arrive in our ecosystem, and the plugin-review question it raises — what capability does installing this grant, and who reviewed it — is the same one the Rank Math story raised last week from a different angle. Skills are becoming an install-time capability grant. We should decide how we feel about that before it ships, not after. ## ✅ Action Items #### DevRel - Developer advocacy: build and publish a `handoff` skill for our own stack using Lipsky's five-part structure; measurable outcome is adoption by 3+ internal teams within a quarter. `#action/h1` - Education: workshop on skill authoring — when a workflow is worth encoding, how to scope it, and how to review one before installing. `#action/h1` - Developer experience: audit which of our repeated workflows are re-explained to a model every session; those are skill candidates by definition. `#action/h1` - Developer marketing: not applicable to this source. `#action/h3` - Community management: prepare guidance on sharing skills safely — never publish one containing a credential, and review borrowed skills before running them. `#action/h1` - Internal/external/cross-project: watch the plugin-ships-skills pattern; it will reach WordPress and the review question arrives with it. `#action/h1` #### OSPO - Policy & Governance: write the policy on AI skills that hold credentials or automate outbound communication. This video demonstrates three such patterns as unremarkable. `#action/h1` - License Compliance: skills borrowed from others (Pocock's Grill Me here) should carry attribution and a license; establish the norm before the ecosystem sets a worse one. `#action/h2` - Security Management: add "credential pasted into model context" to the intake taxonomy. The LMK pattern will be replicated widely and is invisible to current tooling. `#action/h1` - Software Inventory: record installed AI skills as inventory items with their granted capabilities, the same way plugins are recorded. `#action/h1` - Developer Support: office hours on safe automation design — what should never run unattended with write access. `#action/h2` - Community Engagement: not applicable to this source. `#action/h3` - Strategic Alignment: track skill proliferation as a governance surface; it is growing faster than any review process for it. `#action/h2` #### Community - Maintainers: if your project ships a plugin for an AI tool, decide deliberately whether it ships skills and what they are permitted to do. `#action/h1` - Contributors: a shared, reviewed library of skills for common project workflows would lower onboarding cost measurably. `#action/h2` - Sponsors: not applicable to this source. `#action/h3` - Ecosystem: skills are becoming an install-time capability grant across every AI tool. The ecosystem needs a review norm before the install base makes one impossible. `#action/h1` #### WordPress - Core: not applicable to this source. `#action/h3` - Core Performance: not applicable to this source. `#action/h3` - Core AI: directly relevant. If WordPress ships AI abilities that third parties extend, the Remotion bundled-skills model is the design precedent to study — including its review gap. `#action/h1` - Core Program Team: not applicable to this source. `#action/h3` - Mobile: not applicable to this source. `#action/h3` - Playground: Playground is an ideal sandbox for demonstrating what a skill does before granting it real access. `#action/h2` - Community: meetup topic — build your first AI skill, and review someone else's. `#action/h2` - Training: a Learn WordPress lesson on AI skill authoring and safe review would land well and does not exist. `#action/h1` - Marketing: not applicable to this source. `#action/h3` - Support: not applicable to this source. `#action/h3` - Design: not applicable to this source. `#action/h3` - Documentation: document our own repeated workflows in skill-ready form — the specification work is identical. `#action/h2` - Photos: not applicable to this source. `#action/h3` - TV: not applicable to this source. `#action/h3` - Polyglots: not applicable to this source. `#action/h3` - Accessibility: the `visualize` pattern cuts both ways — richer visual output is worse for screen reader users unless the artifact is built accessibly. Worth a position. `#action/h1` - Meta: not applicable to this source. `#action/h3` - CLI: WP-CLI workflows are strong skill candidates; the multi-step chaining pattern maps directly. `#action/h2` - Hosting: not applicable to this source. `#action/h3` - Tide: not applicable to this source. `#action/h3` - Openverse: not applicable to this source. `#action/h3` - Themes: not applicable to this source. `#action/h3` - Plugins: the plugin-ships-skills pattern is the review question of the next two years. Start the conversation now. `#action/h1` #### The WPCC - Programs: not applicable to this source. `#action/h3` - Projects: a reviewed, community-maintained skill library for WordPress workflows would be a concrete, useful project. `#action/h2` - Funding experiments: not applicable to this source. `#action/h3` ### FAIR - Technical independence: the `handoff` pattern is portability infrastructure — it keeps work movable between models and reduces lock-in to any single provider. Worth naming as such. `#action/h1` - Policies and covenants: not applicable to this source. `#action/h3` - Productized governance artifacts: a review checklist for AI skills — what capabilities does it grant, does it hold credentials, does it act unattended — is a small, shippable artifact. `#action/h2` ## 🧪 Hypotheses 1. **Skills improve output consistency more than output quality.** The mechanism is specification stability across invocations, not added capability. Lipsky claims both; only the first follows from how skills work. `#atom/literature/h1` 2. **Conversational skill authoring drives adoption more than any other design choice.** If creation cost exceeds re-explanation cost, no skill gets created. "Describe it in a chat and it saves it" collapses that cost to near zero. `#atom/literature/h1` 3. **Credentials pasted into model context will become a recognized incident class within 18 months.** The LMK pattern — API token into a chat — is being taught as routine, is invisible to current secret-scanning, and scales with skill adoption. `#atom/literature/h1` 4. **Skills that chain browser control, mail write access, and filesystem writes will produce documented unintended-action incidents faster than single-capability skills.** Compound capability, unattended execution, no human in the loop. 5. **Plugin-bundled skills will outnumber user-authored skills within two years.** Remotion ships 12 per install; individuals author a handful. Distribution beats authorship. 6. **Attribution norms for shared skills will not emerge organically.** Lipsky's credit to Pocock is notable precisely because it is unusual; absent a convention, borrowed prompts will circulate unsourced. `#atom/literature/h2` ### 📊 Methodology - **H1 (consistency vs. quality).** Run 30 identical tasks with and without a skill, across 3 task types. Have blind raters score quality on a rubric, and separately compute variance across runs within each condition. Prediction: variance drops substantially with skills, mean quality rises only marginally. This cleanly separates the two claims and is runnable in a day. - **H2 (authoring cost).** Instrument or survey skill creation rates across tools with form-based versus conversational authoring. Confounder to control: tools with conversational authoring are generally newer and more heavily marketed, so cohort by release date. - **H3 (credential incidents).** Track security advisories, provider policy changes, and incident write-ups for the "secret in model context" pattern over 18 months. Pre-register what counts as confirmation to avoid retrofitting. Also testable now: audit your own saved skills and chat history for live tokens. - **H4 (compound capability).** Requires incident data that mostly is not public. Feasible proxy: build a test harness granting a skill each capability alone and then in combination, and count unintended actions across a fixed task battery in a sandbox. This is a real experiment and worth someone running. - **H5 (bundled dominance).** Count skills-per-install across the top 50 plugins in each major AI tool versus median user-authored skill counts from any published telemetry. Simple arithmetic, decisive answer. - **H6 (attribution).** Sample 200 publicly shared skills and code each for attribution presence. Prediction: under 15%. Repeat annually to see whether a norm forms. ### ❓ Questions - What is the actual retention policy for a credential pasted into a chat context, and does it differ across providers and tiers? Nobody teaching this pattern seems to have asked. - Do the terms of service of Grok and Gemini permit automated browser-driven access via a user's authenticated session? The `ask` skill's entire premise rests on the answer. - When a plugin installs 12 skills the user never names, what reviews them? Installing a plugin now grants capability the installer never reads. - How does a skill behave when the underlying model changes? Skills are prose instructions, and prose that steered one model version may steer the next differently. - What is the failure mode of the sponsor-review skill? If Vimeo's layout changes mid-chain, does it stop or improvise — and does it still send the email? - Is `grill me` net-positive on total time? It front-loads interrogation to avoid rework. That is a real tradeoff with a real answer, and neither Pocock nor Lipsky reports one. - Should skills be signed or versioned? They are executable-ish artifacts distributed as text with no provenance. - What does `visualize` produce for a screen reader user? Interactive artifacts are an accessibility regression unless deliberately built otherwise. ### 🔮 Future Research - **Skills as a distribution channel for capability.** The Remotion pattern — install a plugin, receive twelve model-selected skills — is a genuinely new software distribution shape. Its security, versioning, and review properties are unstudied and the install base is growing now. - **Context transfer between heterogeneous agents.** `handoff` is a folk solution to a real problem. A principled study of what must be in a transfer brief for a receiving model to continue work without regression would be broadly useful and is empirically tractable. - **Interrogation-before-execution as a technique.** Does forcing specification up front reduce total task time, or merely relocate it? Measurable with a straightforward controlled study, and the answer would change how a lot of people prompt. - **Secrets in conversational context.** An emerging class of credential exposure that existing scanning does not see, being taught as best practice in tutorial content. The gap between the practice and the tooling is the research opportunity. - **Skill drift across model versions.** Prose instructions tuned against one model, executed by its successor. Whether skills silently degrade at version boundaries is testable and consequential for anyone maintaining a library. ### 💡 Implications - **Practice.** Steal `handoff` and `grill me` today; both are model-agnostic and pay for themselves quickly. Do not steal the LMK credential pattern — route notifications through something that does not require a long-lived token in chat context, or accept the exposure knowingly. - **Policy.** Organizations need a stance on AI skills before employees accumulate libraries of them. The three questions that matter: does it hold a credential, does it act unattended, and does it write anywhere outside the conversation. Any skill answering yes to two of three deserves review. - **Security.** Two novel surfaces are demonstrated here as routine. Credentials in conversation context are invisible to conventional secret scanning. Compound skills — browser plus mail plus filesystem, running on a schedule — are an automation category with no established review practice. Both will scale before controls do. - **Education.** Skill authoring is becoming a core knowledge-work competency, and it is teachable. The specification discipline it demands (state the goal, the constraints, the output shape) is the same discipline that improves technical writing, so the training transfers. - **Funding models.** The bundled-skills pattern shifts leverage toward whoever controls plugin distribution — the same concentration dynamic as app stores and plugin directories, arriving in a new layer with no review infrastructure yet built. ### 🧠 Learning Styles Assessment - **Reading level estimate:** spoken register; Flesch-Kincaid Grade Level approximately 8–9, Flesch Reading Ease roughly 60–68. Short sentences, second person, minimal jargon, and every technical term demonstrated rather than defined. Deliberately accessible. #### 👀 Visual learners - Present: excellent, and the video's core strength. Every skill is shown on screen — the before/after visualize comparison, the Pushover setup clicks, the skills menu, the handoff document, the generated pet. The cold open is a pure visual argument. - Improve: no summary card listing the seven skills, so a viewer cannot capture the set without scrubbing. No architecture diagram for the multi-step skill, which chains four services and would be far clearer as a flow. #### 🔊 Aural learners - Present: clear narration, steady pace, consistent signposting between skills. - Improve: the auto-captions are poor — "chat GBT," "Chad GPT," "Matt PCO," "Grock" — which affects anyone relying on captions, and degrades search and transcript reuse. Human-corrected captions would fix a real accessibility issue, not just a cosmetic one. #### 🗣️ Verbal learners - Present: each skill gets a plain-language purpose statement before the demo, which is the right order. The handoff five-part enumeration is crisply verbalized. - Improve: the Matrix analogy actively misleads about mechanism. "Skills don't give the model new abilities; they stop you re-explaining the ones it has" would be both accurate and more useful. #### 🧩 Kinesthetic learners - Present: the strongest dimension. Copy-paste prompts provided free with no email gate, click-by-click Pushover setup, and an explicit "copy this, paste it, it will walk you through the rest." Viewers can act immediately. - Improve: no verification step — nothing tells you how to confirm a skill installed correctly or is firing. And no teardown instructions: how to revoke the Pushover token, remove a skill, or undo the automations. #### 🧠 Logical learners - Present: sensible progression from simple preloaded skill to compound multi-step automation, ending on the structural observation about bundled skills. That ordering is an argument, not just a list. - Improve: the consistency claim gets one demonstration and is generalized to "every time." No failure cases, no cost discussion, no situations where a skill is the wrong tool. A logical viewer wants the boundary condition and never gets one. #### 🧑‍🤝‍🧑 Social learners - Present: credits Matt Pocock by name; acknowledges the sponsor; invites engagement. - Improve: no community dimension — no mention of where skills are shared, whether others' skills can be trusted, or how a team might maintain a shared library. Skills are framed as a solo practice, which is not how they will actually be used. #### 🧘‍♀️ Solitary learners - Present: self-contained and immediately actionable alone; the free resource page removes the last friction. - Improve: no guidance on evaluating whether a given skill is worth creating, so a solitary viewer may build a library of skills they use once. #### 🌿 Naturalistic learners - Present: nothing environmental. - Improve: scheduled automations running "three or four times per day" on an always-on Mac Mini, each triggering model inference, have a real energy cost that never comes up. Ambient always-on AI is a consumption pattern worth naming. ### 📊 Accessibility Assessment *Assessment of the video as published on YouTube.* #### Screen reader semantics - Present: standard YouTube chrome provides landmarks and controls; the title is descriptive and accurate. - Improve: heavy reliance on unnarrated on-screen action — "click on the left here where it says plugins," "you'll see it right here" — is meaningless without sight. Deictic reference ("here," "this," "right there") is the dominant navigation instruction throughout. Naming the target instead ("in the left sidebar, Plugins") would cost nothing and fix most of it. WCAG 2.2 AA 1.2.3 Audio Description or Media Alternative, 1.3.1 Info and Relationships. #### Keyboard navigation - Present: the YouTube player is keyboard operable by default. - Improve: not applicable to the content itself; the workflows demonstrated are mouse-driven in the ChatGPT desktop app, and no keyboard path is mentioned for any of them. #### Contrast and readability - Present: screen recordings at native resolution, generally legible. - Improve: verify that on-screen text in recorded UI — API tokens, menu labels, the skills list — remains readable at mobile sizes. Small-UI screencasts routinely fail on phones, which is where much of the audience watches. WCAG 2.2 AA 1.4.3 Contrast (Minimum), 1.4.4 Resize Text. #### Text resizing and reflow - Present: YouTube handles player reflow. - Improve: recorded UI does not reflow — text sized for a desktop capture stays that size on a phone. Zooming into the specific region being discussed is the standard fix and is not used here. WCAG 2.2 AA 1.4.10 Reflow. #### Media and captions - Present: auto-captions exist. - Improve: **this is the primary gap.** Auto-captions mangle the central term of the video — "chat GBT," "chatbt," "Chad GPT," "chat GPD" — plus "Matt PCO" for Matt Pocock and "Grock" for Grok. For a Deaf or hard-of-hearing viewer, the product name is unreliable throughout. No audio description track for the substantial unnarrated screen action. WCAG 2.2 AA 1.2.2 Captions (Prerecorded), 1.2.5 Audio Description (Prerecorded). #### Forms and interactions - Present: not applicable to video content. - Improve: the linked resource page should be checked separately for labeled controls and target size, since it is where the actual interaction happens. #### Standards reference - Primary gaps: **1.2.2** (unreliable auto-captions on the core terminology) and **1.2.5 / 1.2.3** (dense unnarrated visual instruction with deictic-only reference). Both are fixable in post at modest cost, and both matter more than usual here because the content is instructional — a viewer who cannot follow the screen cannot complete the task. ### 🌍 DEIB Assessment #### Cultural breadth - Present: the subject matter is tool-generic and not culturally bounded. - Improve: assumes a specific and expensive context — ChatGPT Work (a paid tier), a desktop app, an always-on Mac Mini for automations, subscriptions to Claude, Gemini, and Grok simultaneously, plus a $5 Pushover purchase. That stack costs real money monthly, and the video never acknowledges it as a prerequisite. Viewers on a free tier or a single subscription cannot follow most of this, and are not told so. #### Inclusive language - Present: plain, welcoming, second-person, no gatekeeping vocabulary. Terms are demonstrated rather than assumed. - Improve: "If you're anything like me" and "you've probably never seen it" presume a shared baseline. The Matrix reference assumes familiarity with a 1999 film — mild, but it is the video's central explanatory analogy, so anyone who missed it loses the framing. #### Neurodiversity support - Present: strong. Consistent segment structure (name → problem → demo → invocation), explicit transitions, predictable pacing. The repeated pattern makes the video easy to follow and easy to re-enter after losing attention. - Improve: no chapter markers mentioned and no on-screen summary of the seven, so re-finding a specific skill requires scrubbing. For a reference video people will return to, that is a real cost. #### Representation - Present: credits Matt Pocock specifically and by name for the Grill Me pattern. - Improve: no other voices, and the family budget example ("all of my family's budgeting, savings, and spending plans") presumes a household financial structure that is one common arrangement among several. Minor, but it is the only personal example in the video. #### Psychological safety - Present: encouraging tone throughout, no shaming of viewers who have not adopted these tools, and honest that his own multi-step skill is not directly reusable. - Improve: **the significant gap.** Viewers are instructed to paste an API token into a chat context and to let an automated system send email and modify local files unattended, with no mention that either carries risk. A viewer who follows the instructions exactly and is later harmed was given no basis to make an informed choice. Instructional content that expands a system's authority owes the viewer the tradeoff, not just the steps. ### ✅ Logical Validity #### Strong reasoning - **Problem-first structure for every skill.** Each segment names a concrete recurring irritation before showing the fix — notifications that never arrive, context lost when switching models, the model charging ahead on a half-understood request. Motivating a tool by the pain it removes is sound instructional reasoning, and it lets the viewer self-select which skills matter to them. - **The demonstrated before/after (00:00, 01:47).** The cold open and the visualize segment both show the unskilled output beside the skilled one. Direct comparison on identical input is the right evidence for a claim about output quality, and it is more persuasive than any assertion could be. - **Correct generalization boundary on the multi-step skill (11:45).** *"obviously, this isn't a skill that you're going to be using. However, I want you to think about how you can find any of these multi-step workflows that you repeat very often."* Explicitly separating the non-transferable instance from the transferable pattern is disciplined teaching, and many creators would have oversold the specific artifact. - **Honest attribution (10:48).** Crediting Matt Pocock as the origin of Grill Me, characterizing the difference between the versions, and pointing to both is unprompted intellectual honesty in a genre where borrowed prompts routinely circulate unsourced. - **Non-obvious structural observation (15:29).** Recognizing that bundled plugin skills are model-selected and need not be known by name identifies a genuine shift in how capability is distributed. That is analysis rather than instruction, and it is the most intellectually substantive moment in the video. #### Fallacies - **False analogy** — 00:00: *"It's kind of like that scene in the Matrix, where with one simple upload, Neo was suddenly a kung fu master."* Neo gained a capability he did not have. A skill adds no capability the model lacks; it eliminates re-specification cost and stabilizes output. The analogy misdescribes the mechanism, which matters because it predicts the wrong use cases — viewers will expect skills to make the model able to do things it cannot. **Fix:** "skills don't teach it anything new, they stop you explaining the same thing every time." `#fallacy/false-analogy` - **Absolute overclaim** — 00:52: *"you'll never have to repeat yourself ever again."* Falsified by skill drift, model version changes, evolving workflows, and any task outside the skill's scope. **Fix:** "for workflows you repeat, you'll stop re-explaining them." `#fallacy/unsubstantiated-claim` - **Hasty generalization** — 00:00: *"If you use the right skills, you will get better, more consistent outputs every time you use them."* One demonstrated comparison supports "sometimes noticeably better on suitable tasks," not "better every time." "Consistent" is defensible from the mechanism; "better" is not. **Fix:** separate the two claims, since only one follows from how skills work. `#fallacy/hasty-generalization` - **Suppressed evidence — credentials** — 03:36: the LMK walkthrough instructs viewers to give ChatGPT a Pushover user key and API token, with no mention of retention, scope, rotation, or export risk. Presenting credential exposure as a routine setup step, in content explicitly designed to be copied, propagates the pattern at scale. **Fix:** one sentence naming the tradeoff and how to revoke. `#fallacy/suppressed-evidence` - **Suppressed evidence — terms of service** — 08:57: the `ask` skill drives Grok and Gemini through authenticated browser sessions, justified as *"it takes advantage of the subscriptions I'm already paying for."* Automated UI access via a user session is commonly restricted by providers' terms. The video raises neither the restriction nor the account risk. **Fix:** acknowledge the question and let viewers decide. `#fallacy/suppressed-evidence` - **Suppressed evidence — compound capability** — 11:45: the sponsor-review skill combines browser control, Gmail write access, and local filesystem writes, unattended on a schedule. Presented purely as payoff; the blast radius is never named. **Fix:** state what the chain is authorized to do without a human present. `#fallacy/suppressed-evidence` - **Appeal to novelty** — 14:37: the pets segment is presented as a feature worth covering when it is a completion notification with a costume. Harmless, and it does spend runtime that the security caveats needed. `#fallacy/red-herring` #### Tone of voice Friendly, efficient, and practitioner-credible — a person showing you their actual desk rather than performing expertise. The register is enthusiastic without hype inflation: he says a skill is "one of my favorites" rather than "will change your life," admits his own version of Grill Me is a simplification of someone else's better work, and tells viewers plainly that one of the seven is not reusable by them. The engagement asks (thumbs up, watch to the end, click the next video) are present and unembarrassed but not manipulative, and the sponsor read is clearly bounded and verbally disclosed at both ends. The tonal gap is the uniform positivity of the risk framing. Every skill is presented as pure gain, and the video's most consequential instructions — paste a token, let it send mail, let it write to your vault — are delivered in exactly the same cheerful register as the segment about a paperclip pet. That flatness is the problem: nothing in the delivery signals which two minutes of this video deserve more care than the others. A single change in tone at the credential step would do more for viewer safety than any added disclaimer. ### 🛡️ DARVO Analysis DARVO describes an accused party's response pattern — Deny, Attack, Reverse Victim and Offender. **No DARVO pattern is present.** This is instructional content with no allegation, no accuser, and no defensive posture. The genre does not produce the pattern. Component by component: - **Deny:** Absent. Nothing is denied because nothing is alleged. The relevant absence is *omission* rather than denial — the security, terms-of-service, and compound-capability risks documented above are never mentioned. That is a completeness failure in instructional content, not a denial, and conflating the two would misrepresent what is happening. Every affirmative claim in the video is either demonstrated on screen or plainly true; the problems are overclaims and silences, not falsehoods. - **Attack:** Absent, and notably. No competitor, tool, or person is disparaged. ChatGPT's shortcomings are named plainly — unreliable notifications, no real-time X access, charging ahead on underspecified requests — and treated as problems to route around rather than grievances. Other AI tools are presented as complements with genuine comparative strengths (Grok for X, Gemini for images) rather than rivals. Matt Pocock's version of Grill Me is called "quite intense," which is a characterization offered while crediting him, not a criticism used to elevate the speaker's own fork. - **Reverse victim/offender:** Absent. No grievance is claimed and no persecution narrative appears. The speaker positions himself as a practitioner sharing working methods, and is candid that one of his examples will not transfer to viewers. **Emotional cues and rhetoric.** Consistently upbeat and low-pressure. The persuasive devices in play are the Matrix analogy (which overclaims mechanism), standard engagement solicitation, and a disclosed sponsor read — ordinary creator-economy rhetoric, none of it defensive. **Related fallacies.** The three `#fallacy/suppressed-evidence` findings above capture the real weakness. They describe an instructional gap in content with no adversary — a materially different and less serious problem class than DARVO, and worth stating plainly so the section's vocabulary does not imply otherwise. `#darvo/deny` `#darvo/attack` `#darvo/reverse` — all three assessed. None present. ## 🎯 Conclusion - **Bottom line.** A well-made, concretely demonstrated tour of seven ChatGPT Work skills, whose real value is two transferable patterns rather than the list. **Handoff** — packaging goal, progress, decisions, assets, and next step into a portable brief — is the best idea in the video and generalizes to any pair of agents. **Grill me** — forcing the model to interview you and withhold execution until the spec is settled — is the second, and Lipsky credits Matt Pocock for it unprompted, which is more than the genre's norm. The closing observation about Remotion's twelve auto-selected bundled skills identifies a genuine shift: skills are becoming a distribution format for capability, not just a personal shortcut. The video's failure is one of omission, and it is consistent. It teaches viewers to paste a live API token into a chat context, to have ChatGPT operate other AI services through their authenticated browser sessions, and to run an unattended chain holding browser control, Gmail write access, and filesystem writes — all in the same cheerful register as a segment about a desktop paperclip. Nothing said is false. What is missing is any signal about which instructions deserve more care than others. - **Next steps for viewers.** 1. Build a `handoff` skill this week using the five-part structure — goal, done, decisions, files, next. It is model-agnostic, it pays for itself the first time a project stalls, and it reduces lock-in to any single provider. 2. Try `grill me` on your next underspecified task, and go to [Matt Pocock's original](https://www.mattpocock.com/) before adopting a simplified fork — decide for yourself whether "quite intense" is a bug or the feature. 3. Check whether `/visualize` is already in your install. It ships preloaded, it is the lowest-risk item here, and by his account most people never touch it. 4. Do **not** copy the LMK credential pattern unexamined. Before pasting any token into a chat, know your provider's retention and training policy, and know how to revoke it. Then decide. 5. Before letting a skill hold browser control plus mail write access plus filesystem writes on a schedule, write down what it is authorized to do with no human present. If that sentence makes you uncomfortable, split the chain. 6. Watch the bundled-skills pattern. Installing a plugin now grants capability the installer never reads, and no ecosystem — WordPress included — has a review practice for it yet. ## 🔄 Refactor Appendix — The 7 ChatGPT Work Skills I Use Every Day ### 📚 Literature Atoms (H3) #### [H1] Handoff briefs carry goal, progress, decisions, assets, next step > [!quote] ID:rw_lipsky_h1 > "the goal of the project, what's already been done, the decisions you've made, any files or links that it needs, and exactly what should happen next" **Context & use**: - 07:11 — the handoff skill's payload structure. - The most transferable idea in the video. A general-purpose spec for context transfer between any two agents, and portability infrastructure against provider lock-in. **Tags**: #xref #articleseed #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=Dkdgut6i2xY&t=431) <!-- atomicity: 1 idea ✓ --> #### [H2] Skills stabilize specification rather than add capability > [!quote] ID:rw_lipsky_h2 > "Without skills, you have to reexplain your workflows to chat GPT every time you use it. And that costs you time and gives you inconsistent results." **Context & use**: - 00:00 — the video's accurate framing, immediately undercut by its own Matrix analogy. - The correct mental model, and it predicts which tasks skills help with: repeated, well-specified work you keep re-explaining. Not work the model cannot do. **Tags**: #xref **Highlight**: [view highlight](https://www.youtube.com/watch?v=Dkdgut6i2xY&t=0) <!-- atomicity: 1 idea ✓ --> #### [H3] Plugin-bundled skills are model-selected, not user-invoked > [!quote] ID:rw_lipsky_h3 > "you don't need to know what all these skills are because as you are using the remotion plugin... chat GPT will automatically use the skills it needs" **Context & use**: - 15:29 — Remotion ships 12 skills per install. - Skills as a distribution format for capability. Installing a plugin becomes an install-time capability grant nobody reads — the same review gap the Rank Math Application Password story raised from another direction. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://www.youtube.com/watch?v=Dkdgut6i2xY&t=929) <!-- atomicity: 1 idea ✓ --> #### [H4] Interrogation-before-execution blocks premature building > [!quote] ID:rw_lipsky_h4 > "force chat GPT to interview you first before taking action... it doesn't begin building anything until both you and the AI are on the same page" **Context & use**: - 10:48 — the grill me skill, credited to Matt Pocock. - Front-loads specification to avoid rework. Whether it reduces total time or merely relocates it is an open, testable question neither author reports on. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=Dkdgut6i2xY&t=648) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Portable context briefs make model switching a technique, not a defeat **Claim** (one sentence, Courtney's voice): When a model stalls, transfer a structured brief rather than restarting, so switching models becomes a deliberate move instead of lost work. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds wherever multiple capable agents are available and context is expensive to rebuild — cross-model work, agent handoffs, human-to-agent handoffs, and returning to your own stalled project weeks later. Weakens when the receiving agent lacks access to the referenced files, which is why the brief must name assets rather than assume them. Degenerate case: a brief longer than the work it describes. **Links** (≥1 with annotation): - [[Context engineering]] — because extends - [[Vendor lock-in]] — because applies-in provider independence - [[Incident handoff]] — because contrasts on urgency and completeness tradeoffs <!-- atomicity: 1 idea ✓ --> #### [P2] Installing a plugin is now a capability grant nobody reads **Claim** (one sentence, Courtney's voice): When plugins ship skills the model selects on its own, installation silently grants capability the installer never inspected. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for any ecosystem where extensions bundle agent-invocable instructions — AI tool plugins today, WordPress plugins shortly. Sharpens as the count per install rises (Remotion ships 12) and as selection becomes automatic. Does not apply where extensions expose only APIs a human explicitly calls. Diagnostic: after installing, can you enumerate what the model may now do that it could not before? **Links** (≥1 with annotation): - [[Application Passwords]] — because extends the same install-time-consent gap from another direction - [[Supply chain trust]] — because extends - [[Plugin review]] — because applies-in WordPress `proof_state: working` <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Build a `handoff` skill for my own stack — five-part structure. Highest-value item in this batch. #fleeting #inbox - [F2] Compare Matt Pocock's original grilling skill (I have `mattpocock-skills:grilling` installed) against Lipsky's scaled-down fork. Which register actually fits my work? #fleeting #inbox - [F3] Audit my saved skills and chat history for live credentials. The LMK pattern is being taught as routine and conventional secret scanning cannot see it. #fleeting #inbox - [F4] Remotion ships 12 bundled skills — I have Remotion repo skills (`add-expert`, `writing-docs`). Check what capability those grant that I never reviewed. #fleeting #inbox - [F5] Post idea: "installing a plugin is a capability grant nobody reads" — connects Rank Math Application Passwords, OpenClaw autonomy defaults, and plugin-bundled AI skills into one argument. Three sources, one thesis. #fleeting #articleseed - [F6] Does `/visualize` output survive a screen reader? Interactive artifacts as an accessibility regression is worth testing and writing up. #fleeting #inbox - [F7] Open question worth measuring: does interrogation-before-execution reduce total task time or just relocate it? Nobody reports a number. #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Build a handoff skill using the five-part structure and test it Claude→Codex — tooling #things - [ ] Audit saved skills and chat history for exposed credentials — security #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/The 7 ChatGPT Work Skills I Use Every Day.md|The 7 ChatGPT Work Skills I Use Every Day by Paul J Lipsky]] ## Highlights - kind of like that scene in the Matrix ([View Highlight](https://read.readwise.io/read/01m1c6btv1cr90yy7bd6j5xzg0)) - Note: The framing to resist. Neo GAINED a capability he didn't have; a skill adds nothing the model lacks — it removes re-specification cost and stabilizes output across invocations. The distinction is practical, not pedantic: it predicts what skills actually help with (repeated, well-specified work you keep re-explaining) versus what they won't touch (anything the model can't already do). Same paragraph also carries "you'll never have to repeat yourself ever again," falsified by skill drift and model version changes. - Tags: fallacy/unsubstantiated-claim, fallacy/false-analogy - the decisions you've made, any files or ([View Highlight](https://read.readwise.io/read/01m1c6c15s7hryrs853wmaa2ea)) - Note: Best idea in the video and the one to steal. The handoff payload is five parts: goal · what's already done · decisions made · files and links needed · exactly what happens next. Makes model-switching a deliberate technique rather than lost work, and doubles as portability infrastructure against provider lock-in. Generalizes to any two agents — Claude→Codex, agent→human, or your own stalled project three weeks later. His worked example is honest: iPhone app stalled in ChatGPT, handed to Claude, unblocked, handed back. - Tags: atom/literature/h1, action/h1 - original GrillMe skill was developed by ([View Highlight](https://read.readwise.io/read/01m1c6c44xycs76nj0aytrnmez)) - Note: "Matt PCO" is an ASR error for Matt Pocock — whose grilling skill I already have installed via mattpocock-skills. Unprompted, specific attribution is genuinely unusual in this genre, where borrowed prompts circulate unsourced. Worth testing his original against Lipsky's scaled-down fork: "quite intense" may be the feature rather than the bug, depending on how underspecified the task is. Open question neither author answers: does interrogation-before-execution reduce total task time, or just relocate it? - Tags: atom/literature/h2 - advantage of local file access when it's ([View Highlight](https://read.readwise.io/read/01m1c6c79jj58sah56507zddk4)) - Note: ⚠️ Compound blast radius, presented as pure payoff. The sponsor-review skill holds browser control + Gmail WRITE access + local filesystem writes to an Obsidian vault, running unattended on a schedule. Each capability is defensible alone; chained and automated it describes a system that can send mail as him and modify his notes with no human in the loop. Note the register — this arrives in exactly the same cheerful tone as the desktop-paperclip segment immediately after. Nothing signals which instructions deserve more care than others. - Tags: fallacy/suppressed-evidence - as soon as you install the plug-in. ([View Highlight](https://read.readwise.io/read/01m1c66ack2x3ran9hp01qkjc2)) - Note: Remotion's bundled skills install automatically with the plugin. See the companion highlight on "plugins, you don't need to know what they" for the full argument: skills as a distribution format for capability, and installing a plugin as a capability grant nobody reads. - token, and you also need to give that API ([View Highlight](https://read.readwise.io/read/01m1c6by0c0x0477p0p55w87mh)) - Note: ⚠️ The Pushover setup has you paste a long-lived API token into a chat context, taught as a routine step in content explicitly designed to be copied. No mention of retention, scoping, rotation, or what happens when that context is exported or used for training. This is a credential-exposure class conventional secret scanning cannot see, and it propagates at the scale of the tutorial's reach. ACTION: audit my own saved skills and chat history for live tokens. - Tags: fallacy/suppressed-evidence, action/h1 - plugins, you don't need to know what they ([View Highlight](https://read.readwise.io/read/01m1c6cab2h6xsd4yrzkgcnbch)) - Note: The most structurally interesting moment in the video. Remotion ships 12 skills per install and the model selects among them by itself — so skills become a DISTRIBUTION FORMAT for capability, not a personal shortcut. Installing a plugin is now a capability grant nobody reads. Same install-time-consent gap as the Rank Math Application Password story, arriving from a completely different direction. This is the through-line for a post. - Tags: atom/literature/h1, atom/zettel/z1