Agent Harness Architect

About this skill

The difference between an agent that works and one that flails is the harness — the scaffolding around the model: the system prompt, the tool definitions, context management, guardrails, and the feedback loop. Same model, wildly different results depending on the harness. And the most-neglected piece is the one at the center: the system prompt itself, which teams grow by accretion until it's a contradictory wall of text the model half-follows. Agent Harness and System Prompt Architect designs that core. Describe the agent's job, its tools, and its constraints, and it writes a structured system prompt — role and objective, capabilities with when-and-when-not-to-use-each-tool guidance, boundaries that separate hard nevers from confirm-before-acting, a method, a precise output contract, and defined failure and escalation behavior — written so every instruction is specific, testable, priority-ordered, and free of the contradictions that make models drop rules at random. Then it maps the whole harness: which concerns the prompt should own versus what belongs in tool definitions, context management, guardrails and hooks, or evaluations — because a hard limit stated in prose is a hope, not a control, and format belongs in the output contract, not buried in a paragraph. The download includes three reference files: a system-prompt structure guide, a harness component map, and a worked sample that turns a misbehaving refund agent into a structured harness. It designs the harness and writes the instructions from what you describe — it doesn't run the agent or test the prompt against a model, so pair it with an evaluation to confirm behavior. Works with Claude Code, Cursor, Codex CLI, Gemini CLI, and any SKILL.md agent.