Most important take away

Most people waste enormous amounts of time on AI work because they cram every task into a one-off prompt, when the real leverage in 2026 lives in the scaffolding around the model — prompts, skills, plugins, MCPs/connectors, and hooks/scripts. Learning to disambiguate these layers and package repeatable work into plugins (which can bundle skills, scripts, hooks, and connectors) is what separates people who get 10x value from AI from those who keep re-prompting forever. You do not need to be an engineer to build this scaffolding anymore — domain experts who understand their own workflows are the ones who should be encoding them.

Summary

Actionable insights:

• Stop overloading prompts. If you find yourself pasting the same context, instructions, or process into a chat repeatedly, you are paying a “human plug-in” tax. Convert it into a skill or plugin.
• Use the right tool for the right scope:
  • Prompt = one-off, temporary, highly situational task. Still worth writing well.
  • Skill = a reusable, shareable markdown document that teaches an LLM your house style or repeatable process (PR reviews, marketing docs, structured outbound emails). Portable across Claude, Codex, etc.
  • Plugin = a packaged workflow that can bundle skills, scripts, hooks, MCP servers, connectors, assets, and commands under one installable name. Built for sharing across a team.
  • MCP / app connector = the “universal plug” that lets agents reach live data and systems (Salesforce, Slack, Figma, GitHub). A plugin can contain an MCP, but is bigger than one.
  • Hooks / scripts = deterministic checks (formatters, schema validators, tests, JSON contract checks, pre-stop reviews). Do not let the model “imagine” running these — actually run them.
• Apply the 20/80 rule to skills: 20% of your skills will deliver 80% of value. Find the high-frequency, high-sensitivity workflows first.
• Draw clean edges around workflows. Splitting “customer success” into separate plugins (refunds, activation, upgrades) is usually better than one mega-plugin. A workflow should have one job.
• Treat these layers as Lego bricks, not competing tools. They compose; they are not in competition.
• Build proactively — do not wait for OpenAI or Anthropic to ship the plugin you need. Claude’s “cloud design” is essentially a plugin that became a product, which signals where the platforms are heading.
• Concrete plugin candidates: editorial first-pass review, weekly business reports (spreadsheets + Slack + docs + dashboards), customer briefs (email + CRM), hiring readouts, design pulls from Figma.

Career advice:

• The most valuable emerging skill is workflow decomposition: looking at work and identifying the right boundaries to package into a plugin. Very few people can do this well, and it is “gold” right now.
• Domain expertise from non-technical roles is the differentiator in 2026. The people who know what good output looks like, which sources matter, and which steps get forgotten are the ones who should be authoring scaffolding — not waiting for engineers.
• Educate your leadership. Many C-suite leaders (outside the CTO) do not understand the harness, which leads to bad decisions like “why aren’t you using AI?” when you actually need a deterministic script. Share this mental model upward to get organizational support for AI transformation.
• Invest time in starter templates. Customizing an existing plugin is far easier than starting from a blank page — keep a personal library.
• Keep human judgment in the loop. The goal is not to plugin-ify everything; some work should stay prompts, some should stay scripts, and final judgment should stay human.

Chapter Summaries

1. The mech-suit metaphor — LLMs alone are not the story; the scaffolding (prompts, skills, plugins, MCPs, hooks, scripts) is what makes agents capable. The real product question is how to build that scaffolding yourself.
2. Why this matters now — GPT 5.5 and similar models are good at “messy, multi-part work,” but most people stall because they do not understand the harness, not because they lack tools.
3. Prompts — Best for one-off, temporary, situational tasks. Do not carry permissions, tools, or workflow. Most people over-index here and waste hours.
4. Skills — Reusable markdown documents that encode a process (e.g., house style for outbound emails). Portable across LLMs. Apply a 20/80 lens because people tend to write too many.
5. Plugins — The bigger package. Bundles skills, MCPs, hooks, scripts, assets, and commands into one installable workflow. Shareable across teams. Replaces the “human plug-in” pattern of copy-paste-reason-fetch-check.
6. MCPs and connectors — The “universal plug” to live data and external systems. Often confused with plugins; an MCP can live inside a plugin but is narrower in scope.
7. Hooks and scripts — Deterministic operations (formatting, validation, tests, structural checks, pre-stop reviews) that should never be left to model judgment. They nest inside plugins.
8. Plugins are workflow packaging, not App Store add-ons — Ask “what part of my work has repeatable structure?” rather than “what can I install?” Lego-brick mental model.
9. Drawing workflow boundaries — A high-value, scarce skill: identifying the right unit of work for a plugin. One workflow = one job; split big domains (like customer success) into multiple plugins.
10. Scaffolded vs. unscaffolded agents — Same underlying model, but a scaffolded agent delivers ~10x more value. Humans are a key part of making AI smarter by building the harness.
11. Non-engineers can do this in 2026 — Examples: an editor’s first-pass review plugin, Figma-connected design plugins. Claude’s “cloud design” is essentially a fancy plugin made into a product.
12. Disambiguation recap — Once = prompt; repeated = skill; needs to travel with tools/assets = plugin; needs external system access = MCP/connector; needs verification = hook/script.
13. Educating leadership — Many senior leaders (especially non-CTO C-suite) do not understand this layer, which blocks AI transformation. Share the mental model so the organization can support the work.
14. Closing — Not everything needs to be a plugin; the goal is recognizing repeated, valuable, structured work and packaging it appropriately. This is where the real leverage lives in 2026.