Notion Custom Agents vs Zapier Agents: Which AI Agent Platform Should You Use?
Notion Custom Agents and Zapier Agents take fundamentally different approaches to AI automation. Here's where each one wins — and where it falls short.
Skills > Prompts: How I Teach AI to Work My Way
How reusable AI skills across Claude, Notion, and Linear have replaced prompt engineering as the highest-leverage way to make AI tools actually useful.
We spent 2024 learning to write better prompts. In 2026, the leverage has moved somewhere else entirely. These days, the biggest gains I'm seeing, in my own work and across the tools I use, come from giving AI agents structured background knowledge about how you actually work. Not what you want right now, but how you always want it done and how you want things done across your team. The industry is calling these "skills," and they're showing up everywhere.
A skill is a reusable set of instructions that teaches an AI agent how to do a specific task the way you want it done. Think of it like the scene in The Matrix where Neo gets kung fu uploaded directly into his brain — except instead of martial arts, it's "how to draft a project proposal for a work.flowers client" or "how to scaffold a Zapier integration using the CLI."
There were earlier attempts at this. OpenAI's Custom GPTs let you create specialised chatbots with their own instructions, knowledge files, and tools. The problem was that you had to remember which GPT to talk to for which task. If you were working on something that crossed multiple domains — say, turning meeting notes into Linear tickets and then drafting a client follow-up email — you'd need to get one Custom GPT to summarise your conversation, copy the output, open a different Custom GPT, paste it in, and pick up from there. Manually. Like an animal.
Anthropic's approach with Claude skills changed this. Instead of siloed chatbots, skills are reference documents that sit in your project or codebase. Claude reads them at runtime and figures out which skill is relevant based on the conversation. You don't have to remember “who” to talk to; you’re just talking to Claude, and Claude loads the right instructions when they're needed. This is more memory-efficient, too, because the model doesn't need to hold every skill's instructions in context all the time — it effectively “learns” new skills on-demand.
And it turns out this matters a lot. A recent benchmark study, SkillsBench, tested 7 agent-model configurations across 7,308 task runs and found that curated skills raised average pass rates by 16.2 percentage points. The gains varied wildly by domain — from +4.5pp in software engineering to +51.9pp in healthcare — but the direction was consistent. Two other findings stood out: self-generated skills — where the AI writes its own instructions — provided no benefit on average (–1.3pp), and focused skills with 2–3 modules consistently outperformed comprehensive documentation. The skills I've written that work best match this exactly: narrow, specific, and far from encyclopaedic.
How I use skills in Claude
My most-used Claude skill is one I wrote for building custom Zapier integrations using the Platform CLI. A lot of the apps my clients use either don't have a native Zapier integration, or the native one doesn't support the specific trigger or action I need. So I build custom ones.
A couple of months ago, I started using Claude Code with a custom skill that teaches Claude how I want Zapier integrations built, based on the official Platform CLI docs and my own workflow preferences. The difference was immediate. What used to take half a day now takes about 10 minutes.
The most recent example: I discovered AgentMail, a new API-first email platform designed for AI agent workflows. No native Zapier integration. So I pointed Claude at the API docs, and it scaffolded the entire integration, wrote and ran the tests, and deployed it to Zapier. My total hands-on time was about 5 minutes of checking in. The skill is open source on GitHub if you want to try it yourself.
This is what Casey Newton describes as Claude Code "waking people up to LLMs' power to generate tools." The bottleneck between "this app has a public API" and "I can automate workflows with it in Zapier" is gone.
How I use skills in Notion
Notion is where I spend most of my working day, and skills have become central to how I use it.
I built a database of Agent Skills that's shared across my team. Each skill is a page with a description of when it should be used, and the full instructions for how to execute the task. I have skills for drafting project proposals, writing SOWs, generating invoices, creating blog posts, reviewing sales calls, and more. I even have one that helps me create new skills.
The trick is in how my AI instructions page references this database. I have an inline view of the Agent Skills database filtered to skills I'm tagged in. My instructions tell the AI to scan the Description property of every skill first — not the full page, just the description. If a skill's description matches the current task, then the AI opens and reads the full skill page before responding.
skill-creator skill.This is directly inspired by how Claude handles skills: load only what's relevant at runtime, don't try to hold everything in memory at once. It keeps the context window lean and the outputs focused. And because the database is shared, my team members can use the same skills from their own AI instructions pages, or build new ones that everyone benefits from.
Notion recently launched a native skills feature that formalises this pattern inside the product. I've just gotten access and am still finding my way around, but the direction is right: skills as first-class objects that the AI can discover and apply automatically.
For now, I'm keeping my database-driven approach running in parallel while I evaluate the native version.
Skills are showing up everywhere
This pattern is spreading fast. Just this week, Linear launched a major overhaul of their in-app Agent, with skills as a core feature. Their CEO Karri Saarinen framed it this way: as agents absorb more procedural work, the bottleneck shifts from execution to judgment. Skills are how you codify that judgment so it compounds over time rather than getting re-explained in every conversation.
Linear's implementation lets you save a workflow that worked well as a reusable skill, trigger it manually via slash commands, or let Linear Agent automatically apply it when it thinks it's relevant. The use cases they highlight — catching up on projects, drafting issues from meeting notes, synthesising customer feedback — are exactly the kind of recurring workflows where skills shine.
The convergence is telling. Anthropic built skills into Claude. Notion is building native skills. Linear just shipped skills. The pattern keeps repeating because the underlying insight is the same: AI agents become dramatically more useful when you give them structured, human-written procedural knowledge about how you actually work.
What makes a good skill
After building a couple of dozen skills across Claude, Notion, and now experimenting with Linear, a few principles have emerged:
- Keep them narrow. A skill that tries to cover every edge case will perform worse than one that handles a specific workflow well. The SkillsBench data backs this up: 2–3 focused modules beat comprehensive documentation every time.
- Write them from experience, not theory. The same study found that self-generated skills (where the AI writes its own instructions) provided zero benefit. The skills that work are the ones distilled from real workflows you've actually done or specialist knowledge that you uniquely provide.
- Include at least one concrete example. Step-by-step instructions are good. Step-by-step instructions with a worked example of what the output should look like are significantly better.
- Use descriptions as routing logic. Whether it's Claude scanning your
AGENTS.mdfiles or Notion scanning your skill descriptions, the trigger text matters. Write the description as a clear statement of when the skill should activate, not a vague summary of what it does.
- Treat skills as living documents. Every time a skill produces a result that needs manual correction, update the skill. The compounding effect of this is enormous — each correction makes every future use of that skill better.
From the SkillsBench paper: Curated skills raised average pass rates by 16.2pp across 7,308 task runs. but self-generated skills provided no benefit on average (–1.3pp).
Models can't reliably write their own procedural knowledge.
The skills that work are the ones humans write from real experience, and focused skills with 2–3 modules consistently outperformed comprehensive documentation. Less really is more.
The bottom line
The gap between "AI that's impressive in a demo" and "AI that's useful in my actual work" has always been context. Skills are the most practical way I've found to close that gap. They're lightweight to create, easy to share across a team, and the research suggests they can make a meaningful difference when done right — even letting smaller models with good skills match larger models without them.
If you're using AI tools in your work and haven't started building skills yet, pick one task you do repeatedly and write the instructions for it. Be specific. Include an example. See what happens.
And if you want help building a skills library for your team — whether in Claude, Notion, or across your whole tool stack — let's have a chat.
Related Posts
How I Built a Content System That Doesn't Sound Like AI Slop
How I built a content system using Notion, Zapier, Readwise, and AI agents that produces original writing at scale — without sounding like everyone else.
5 Notion Custom Agents Use Cases to Automate Your Workspace
Five practical Notion Custom Agent use cases we've been building—from automated meeting prep to turning Slack chaos into trackable tasks.