Most people meet AI as a chatbot. The outcomes come from what is built around it.
Ask ten people what AI is and most will describe a window where you type a question. That is the part you see. Behind it sit five layers, each one resting on the one before it. Put them together in the right order and an assistant becomes something that delivers outcomes on its own. I call the build order SPARK.
Five layers, in build order
- Standards. The governance and context the agent works within: who it serves, what it can see, and the rules it follows. Every other layer inherits from this one.
- Pipelines. Access to the right data, and the ability to take action in the apps your team already uses. Without them, the agent can only talk.
- Abilities. Teaching it how you want things done: your steps, your format, your standard of good. This is where your know-how lives.
- Runner. The agent itself. It acts on the data, using the abilities, to get the work done.
- Kickoff. What starts the runner: a schedule, or a person asking.
What each platform calls it
Claude and ChatGPT build the same five layers. Only the names differ.
| SPARK | Claude | ChatGPT |
|---|---|---|
| SStandards | Project | Project |
| PPipelines | Connectors | Apps |
| AAbilities | Skills | Skills |
| RRunner | Agent (Claude) | ChatGPT Work |
| KKickoff | Scheduled tasks | Scheduled Tasks |
Product names as of November 2026.
Three things worth knowing
- Bundles have a name. Both platforms call them plugins: a package of skills, apps or connectors, and instructions that installs in one go.
- Skills travel. Claude and ChatGPT use the same skill format, so the work you put into teaching the “how” is not locked to one platform.
- Watch the gap. In ChatGPT, a scheduled task made inside a project can't see that project's files. Test it before you rely on it.
Why the order matters
Most AI pilots that stall have skipped a layer. A runner with no standards guesses at what you want. An agent with no pipelines has nothing to work on. Abilities taught to an agent that can't reach your data are advice nobody can use. And without a kickoff, the whole thing waits for someone to remember it exists.
Start with standards and build forward. Each layer makes the next one worth having.
Every outcome starts with a SPARK.
Which layer is missing in yours?