The Agent Operator’s Manual

Description

The Agent Operator’s Manual — Write Instructions That Make AI Agents Actually Work Across Any Tool

You let the agent do it — Claude, Cursor, Cline, Aider, whatever you’re running this week. Some days the output is perfect. Other days it produces something you wouldn’t put your name on. You re-prompt. You add caveats. You paste in more context. You try a better model. You switch tools. The problem follows you.

The problem isn’t the model. The problem isn’t the tool. The problem is the file that tells the agent what to do.

Most founders are tuning the wrong variable. The model is responsible for roughly 20 percent of agent quality. The instruction file is responsible for the other 80. This book is about the 80 — and the craft transfers cleanly across every agent tool on the market.

Inside this book:

• The Agent Instruction Model (AIM). A six-layer framework — Mission, Tools, Decisions, Output, Error Handling, Quality Gates — wrapped in a Context layer. Every effective instruction file Pat has ever written has these layers, whether the target is a CLAUDE.md, a Cursor rules file, a Cline config, or a custom system prompt. Every failure he has diagnosed traces back to a missing or underdeveloped layer.
• Eight named principles you can apply to any agent on any tool, any frontier model: Scope Boundary, Confidence Threshold, Canary-Pilot-Full, Inspector Independence, Fallback Chain, Debug Loop, Autonomy Spectrum, and the 80/20 Rule of Agent Quality.
• The instruction file from the morning seventy-two posts went live. The full annotated CLAUDE.md, the broken version, the rewrite, and the line-by-line diff that explains why the rewrite would have prevented the crisis. Plus complete starter files for the Writer, the Optimizer, and the Site Monitor — ready to fork into whatever tool you run.

Part of The Sovereign Entrepreneur series — receipts not theory, from a solopreneur running a real operation.