Back to Blog
Agentic AI

Prompts Are Not the Product: System Instructions Inside a Claude Harness

March 22, 202510 min read
PromptsClaudeStructured OutputEvals
Prompts Are Not the Product: System Instructions Inside a Claude Harness

System rules live in the harness

Tone, forbidden actions, and output shape belong in a versioned system prompt. User messages carry the task. InferentCloud stores prompts next to tools and evals so a change is reviewable.

Structured output

Ask Claude for schema-bound JSON, then validate before any tool write. Failed validation retries with a repair instruction. Unchecked free text is not an API.

Evals over vibes

Golden inputs in CI catch prompt drift. Track business metrics (resolution, cycle time) as well as model scores.

Summary

Prompt engineering for an agency is harness engineering: instructions, schemas, retries, and evals around Claude.

Quick Share:WhatsAppLinkedInX