Most of the AI conversation in 2026 happens at the model layer, which model is best, which is cheapest, who shipped the new capability this week. That conversation is interesting and almost entirely useless to a business deciding what to actually do with AI.
The conversation that matters is one layer up: how AI capability gets composed, deployed, and integrated into a business’s actual operations. In the Anthropic ecosystem specifically, the answer to that question increasingly runs through plugins, skills, and the Model Context Protocol (MCP). I’ve been working inside this ecosystem for the last several months, building, customizing, deploying, and I’ve learned things that are worth writing down for anyone considering the same path.
This piece is a practitioner’s-eye view, not a marketing piece. I’m not releasing code here. I’m sharing what I’ve found about what works, what doesn’t, and where the leverage hides.
What plugins, skills, and MCP actually are
A quick orientation, because the vocabulary is moving fast.
Skills are reusable bundles of instructions, scripts, and reference material that Claude can pull in to handle a specific kind of task. A skill is the difference between asking Claude to write a Word document and giving Claude a kit of best-practice-driven Word document templates plus the methodology for assembling them. Skills concentrate institutional knowledge.
Plugins are bundles of skills, MCP servers, slash commands, and configuration that ship together as a single installable unit. Plugins are how a domain, sales, finance, design, and customer support get packaged so that any user can install the whole capability with one command.
MCP (Model Context Protocol) servers are the connective tissue between Claude and external systems. An MCP server gives Claude access to a specific tool, your calendar, your CRM, your project tracker, and your database in a structured way that the model can reason about. MCP is what lets Claude do work in your real systems instead of just talk about doing it.
These three layers compose. A finance plugin might bundle a journal-entry skill, a reconciliation skill, an MCP server connecting to your accounting system, and a slash command that triggers month-end close. The user installs the plugin, types /close, and the system runs.
That’s the architectural picture. Now the lessons.
Lesson 1: The hard part is workflow design, not capability
Once you spend time in the Anthropic ecosystem, the model itself stops being the interesting part. Claude is capable enough that the bottleneck shifts to workflow design, what triggers what, in what order, with what handoffs, against what definition of success.
This maps to a long-running theme in product management: when the technology becomes good enough, the differentiator moves to how the technology is integrated into the surrounding work. Smartphones in 2010 were a hardware competition. Smartphones in 2020 were a software ecosystem competition. Smartphones in 2026 are an AI integration competition. AI capability is following the same trajectory faster.
For businesses adopting AI, this means the question “which model should we use” is the wrong first question. The first question is “what workflow do we want to redesign, and what would good look like for that redesigned workflow.” The model choice falls out of that.
Lesson 2: Skills are how institutional knowledge gets durable
The most undervalued construct in the ecosystem, in my view, is the skill. A well-written skill captures the way a specific kind of work should be done, the steps, the gotchas, the artifacts to produce, the failure modes to avoid, and makes that knowledge invocable by anyone, anywhere, with any model.
The implication for a business: a skill is the closest thing AI gives you to durable institutional knowledge that doesn’t walk out the door when an employee leaves. Codify how your best customer-support rep handles escalations. Codify how your best salesperson qualifies a deal. Codify how your accounting team closes the books. Each of those is a skill. Each one is now reproducible at any scale, on any day, with any team member who pulls the right plugin.
Most businesses have no idea this is possible. They are still thinking of AI as a chat tool. The businesses that figure out the skill model first will compound an advantage that’s hard to catch up to.
Lesson 3: The MCP layer is where vendor decisions get sticky
Every MCP server is a new dependency. They don’t all have the same maturity, the same reliability, or the same security posture. The choice of which MCP servers to install, and which to write yourself, is one of the most consequential architectural decisions in an AI deployment, and it’s mostly being made on autopilot.
Things I look for when evaluating an MCP server:
- Who maintains it (the vendor, a third party, a hobbyist)?
- How are credentials handled? Are there scoped permissions, or is it all-or-nothing?
- What’s the rate-limiting behavior? Can it be configured per workspace?
- What’s the audit trail? Can I see what the model did, or only what it said it did?
- What’s the failure mode when the underlying service is degraded?
Most MCP servers I’ve evaluated answer poorly to two or more of those questions. That doesn’t mean don’t use them, it means use them with eyes open, and design fallback paths for when they misbehave.
Lesson 4: Plugins are how consultants ship product
The plugin format unlocks something genuinely new for consulting. Historically, a consultant produces a deck, a process map, maybe a working spreadsheet, and the engagement ends with handoff. The client now owns a static artifact and is responsible for keeping it alive.
A plugin is different. A plugin is shippable, installable software that captures the consultant’s recommendations as runnable capability. The client doesn’t get a deck, they get a system that runs.
This changes the consulting business model. The deliverable is no longer recommendations. The deliverable is a deployed capability. The pricing follows: a one-time fee for the engagement, then a maintenance relationship for keeping the plugin current as the model and ecosystem evolve.
I think this is one of the most underrated developments of the year. Anthropic has effectively given consultants a packaging format for their expertise that can compound across clients. The consultants who recognize this early will build moats. The ones who don’t will keep delivering decks.
Lesson 5: Customization is where the business value lives
Out-of-the-box plugins are useful but generic. The value extracts when a plugin gets customized to a specific organization’s tools, workflows, and constraints, its actual CRM, its actual project tracker, its actual brand voice, its actual SLA bar.
Customization is where I’ve spent most of my time. It’s where the business value is, and it’s the work that’s hard to do without product judgment. Choosing which skills to retain, which to tune, which to replace; choosing which connectors to wire up; tuning prompts to match the organization’s tone; setting evaluation thresholds for when to escalate to a human, these are product decisions, not engineering decisions, and they don’t happen by accident.
The work I do for clients increasingly looks like this: take a stock plugin, customize it to the client’s environment, deploy it, instrument it, and tune it over the first ninety days against actual usage data. It’s product management applied to AI capability, packaged for a buyer who can’t yet do this themselves.
Concrete examples from my own portfolio
A few patterns that have come out of my own builds and may be useful as reference points:
- AutoBlog is essentially a Claude-orchestrated content workflow that publishes directly into WordPress. The LLM-generation half is straightforward; the value is in the workflow design — site-specific voice configuration, two-pass generation (outline then body), structured output that maps cleanly to Gutenberg blocks, and direct API publish so the human is removed from the friction layer.
- AutoWebDev takes the same orchestration pattern further: instead of just generating copy, it generates a full site specification (architecture, copy, layout, imagery, brand profile) as a reviewable intermediate artifact before any CMS is touched. The site spec becomes the source of truth; the CMS is just the destination. That separation is something most “AI website generators” don’t do, and it’s why those generators feel locked-in while AutoWebDev feels portable.
The common thread across all three: AI capability composed at the workflow layer, with the model treated as one component in a deliberate product architecture, not as the product itself.
Lesson 6: The ecosystem is moving fast, design for change
Six months ago, the plugin format didn’t exist. MCP existed but in early form. Skills were a different abstraction. The ecosystem in six more months will look different again.
The implication: don’t lock into architectures that assume the current state of the ecosystem is permanent. Build with abstractions you can swap. Document your decisions with dates and rationale, so you can re-evaluate without having to reverse-engineer your past thinking. Subscribe to the official changelogs, not the LinkedIn summaries, because the LinkedIn summaries lag.
This is the operational discipline I keep returning to: AI work is not a snapshot, it is a stream. The systems that work are the ones designed to evolve, not the ones designed to be finished.
What this means for businesses
If you’re an SMB or mid-market business looking at the Anthropic ecosystem and trying to figure out where to start, here’s the short version of what I’ve learned.
Start with one workflow, one plugin. Pick a workflow where the cost of a mistake is recoverable and the volume is high enough to matter. Install or build a plugin tuned to that workflow. Run it for ninety days. Measure.
Treat skill development as institutional-knowledge capture. Write skills the way you’d write SOPs, but with the assumption that they’ll be invoked by an AI model rather than read by a new hire. The discipline is similar; the format is different.
Be deliberate about MCP server choices. Each one is a dependency you’re taking on. Evaluate maintainers, credentials, rate limits, audit trails, and failure modes before you wire them in.
Hire help that has shipped in this ecosystem. Not a consultant who has read about plugins. A consultant who has built, deployed, and tuned them, ideally for businesses similar to yours.
If you want help with any of that, that’s exactly what I do at Local Value Marketing. The plugin ecosystem is a real differentiator, and almost no consultant has shipped meaningful work in it yet. The window for businesses to gain compounding advantage is open right now.
Related: Why Most SMBs Are Doing AI Wrong and Building Production AI as a Solo PM.
Rob Lewis is the founder of Local Value LLC and an AI Product Consultant for SMB and mid-market businesses. He has three decades of senior product leadership experience and ships production AI applications. Reach him at [email protected].
