Module 2 · Optimising and operating prompts · Lesson 5
Optimising tools, instructions and skills
An agent's behaviour depends on far more text than its system prompt - tool names and descriptions, error messages, skill descriptions and bodies. How to evaluate each surface (tool selection, skill triggering, end-to-end tasks) and optimise it by hand, with Claude, and with GEPA- and ACE-style methods.
By the end of this lesson you can
- Inventory the full text surface of an agent and say which parts are always in context and which load on demand
- Build tool-selection and skill-triggering evals, and compute precision, recall and F1 for a skill description
- Rewrite tool and skill descriptions using current guidance, and test the rewrite on held-out queries
- Decide when to move system-prompt content into skills or tools, and measure the effect on the agent's task eval
- Explain how GEPA and ACE extend automatic optimisation from single prompts to tools, skills and evolving playbooks