I’m experimenting with a different way of working with Cursor.
Instead of writing a detailed specification, I gave the agent only a few high-level categories for an existing system with a substantial codebase and existing infrastructure.
The interesting part was that Cursor didn’t just generate code. It inferred a useful structure from the existing context: navigation, UI sections, content organization, links, and how the new layer should connect to what was already there.
So the question I’m trying to explore is not “Can an AI build a website from four words?”
It’s:
As the agent’s understanding of an existing codebase increases, how much less explicit specification can we give it while still getting a useful result?
In other words:
minimal intent + rich codebase context → how much specification can the agent infer on its own?
I’d be interested in hearing from people who have tested this on large existing systems, especially across Cursor, Codex and Claude Code.
Personally, I’m finding the less context I give an agent, the better the results. There are exceptions to be sure, but my default these days is to let the agent explore the existing codebase.
The stronger the model, the more the marginal value of my spec goes negative.
In other words, as the agent’s understanding of an existing codebase increases, we can provide less explicit specification while still achieving useful results.
I think “less context” is slightly misleading here.
The context didn’t disappear — the agent retrieved it.
If I give the agent a minimal instruction and it explores hundreds of files, follows imports, discovers architectural conventions, reads project rules, runs the system and inspects the results, that is not a context-free workflow.
It is a shift from human-supplied context to agent-discovered context.
And that distinction matters.
A stronger model may reduce the value of a long specification, but that does not necessarily reduce the value of context. In fact, its ability to identify and retrieve the right context may be exactly why the shorter specification works.
So I’d frame the hypothesis differently:
Model capability ↑ → required explicit specification ↓
but
successful autonomous context discovery ↑
The interesting question is not how little context we can give the agent.
It’s how little intent we can provide before the agent can no longer reliably reconstruct the missing specification from the system itself.