Context-engineering

Post two in the series. How the Legal Framework is actually built in Claude Code: three tiers, three Python scripts, a reflective hook, and the operating loop. Includes the skill as a download, plus the parts that are not right yet.

Our context mechanisms are good at prescription - the rules an agent is told to follow. They are poor at experience - the hard-won nudges we supply at the moment of ambiguity. This is a framework that borrows from our legal system to give that experience a home, the Common Law, and recalls it just-in-time rather than in advance.

The agent space is in the same moment the web was before patterns existed to stabilise it. I think the principle that settles it already has a name.

Most agent teams are org charts with extra steps. The useful distinction isn't what agents do, it's how they think.

Most agent instructions read like mission statements - aspirational, comprehensive, and ignored in practice. After three iterations of building agent teams, I found what actually changes agent behaviour: directives that are specific enough to violate.