The Escape Velocity Manifesto: Beyond the Gravity of Chat

 

The Gravitational Pull

Most of the industry is converging on the same spot: the chat window. Engineers and managers are settling into a loop of prompting, correcting, and re-prompting. The more advanced are using Copilot inside the IDE. A smaller group knows what MCP is, and an even smaller group actually uses it.

It feels productive because it’s visible. You’re typing, the model is answering, something is happening. But treating AI as a chatbot has a low ceiling: it’s slow, it hallucinates under weak context, and it doesn’t scale to the kind of work enterprise teams actually do.

I think of this as the gravity of the average. It’s not wrong, it’s just where the floor is.

Breaking Free

To lead a team well in this shift, I didn’t need to prompt faster. I needed a different operating model.

So I stopped optimizing prompts and started engineering context.

The Methodology: Specs & Skills

The framework I’ve landed on rests on two pieces:

1. Specs are the trajectory. Instead of letting the model infer intent from a paragraph, I give it explicit business rules, architectural constraints, and acceptance criteria up front. The plan is reviewed by a human, then it becomes the contract the implementation has to satisfy.
2. Skills are the engine. Modular, reusable patterns the model can execute the same way every time: PR creation, ticket transitions, test scaffolding. Deterministic where determinism matters, so the creative work happens where it should.

When the context is structured this way, a lot of the friction disappears. Hallucinations drop because the model isn’t filling in gaps it shouldn’t be filling. The system understands the project, not just the last twenty messages. In my team, this workflow has roughly doubled delivery throughput compared to our previous Copilot-in-IDE baseline, without removing human review from the decisions that matter.

The View from Orbit

This isn’t really about coding efficiency. It’s about where engineering leadership spends its attention.

Debating which prompt phrasing works best for a single function is a tax. Architecting the context that makes thousands of those prompts unnecessary is leverage. The shift is from using AI to engineering the system AI operates inside.

That’s the work I think matters now: building the specs, the skills, and the workflows that turn AI from a clever assistant into infrastructure your team can rely on.

Stop fighting gravity. Start engineering context.