The Liminal Prompt
How your agent might be working against you
My frustrations had been steadily building with agent-based test writing for weeks. I’m particular about tests, they’re critical. They’re easy to get wrong. It takes experience to get them right. Despite my extensive skill library Claude continually wrote poor tests. Tests would reach through to underlying modules testing deep abstractions. They were extremely granular, instead of three tests I might wind up with nine. And Claude loved to test variations on a theme, especially when text was involved, that had no real impact on the stability of the underlying code.
I'd glanced at agent prompts before but never stopped to deeply interrogate one. After a frustrating session with a subagent test writer, I hit Ctrl+o and read the prompt it was given. There it was. Clear as day. Claude was working against my own well-developed skills, strongly conflicting with my preferences. Annoying.
My skill library is extensive. On top of that I have agents that use those skills and coordinate with each other. The problem wasn't the library or the agents. It was the hand-off. The primary agent, the orchestrator, was providing too much context when spinning up subagents, and that context was working against the skills and directives I'd carefully built.
These skills capture behavioural- as well as task-oriented perspectives. They understand personas, software flows, even user stories. The orchestrator wasn’t accounting for that. My library subagents already know the how, they just need to be told the what so they can get started.
With a mature skill library, the orchestrator’s job needs to evolve in a specific way alongside it. Especially as you begin encoding task-specific instructions in skills and subagents, it needs to defer rather than deeply instruct. Currently I’m solving for that deferral with a dedicated agent-briefing skill. This lets the primary agent orchestrate a clean hand-off without on-the-fly re-inventing my skill library.
How mature is your agent/skill library? If you’re consistently handing off to subagents, scrutinizing the orchestrator’s prompt is worthwhile. You might find its prompting is undermining your expectations, generating output counter to your own skill library. Annoying.