Signposts

What’s broken. Instructions don’t hold. You tell an AI coding agent the rule, you put it in the prompt, you put it in the file that gets injected at the start of every conversation, and it complies for a while… then quietly drifts back. The longer a session runs, the more the early instructions fade. This is context drift, one of the load-bearing problems of building with AI: if the model can slip a mistake in at any point, you have to re-read everything it writes, every time, just in case. That dread is most of what people mean when they say agentic coding is more trouble than it’s worth.

What I’m building. Signposts is the harness I’ve been building for the last six months: the scaffolding around the model that keeps it on the rails so I don’t have to watch it. The realisation underneath it is that the thing you build to get an AI to do what you want is not a prompt, it’s enforcement. It has two halves:

• Rules. Hard, enforced checks. A new failure mode becomes a new rule, fired the moment the model edits a file (and again before every commit). A rule doesn’t ask, it blocks, and the message it sends back explains itself, so the model reads it, remembers, and gets it right again. A banned pattern simply cannot make it into the codebase. Not once.
• Signposts. Soft, advisory guidance, keyed to wherever you’re working. Where a rule enforces, a signpost shapes: the right pattern for this area, surfaced the moment you touch it, before you go wrong.

So now when I hit a problem I don’t reach for a better prompt, I firm up the harness. You set the goal, then you set the rules that make any other outcome impossible. And once the rails are there, the part I find genuinely exciting: a smaller, cheaper, more efficient model under a strong harness will quietly beat a big model run carelessly. The harness, not the model, becomes the biggest lever you have.

It’s early as something anyone else could pick up. Right now it’s the harness running this very site, and the work ahead is turning it into something you could lift into a project of your own.