# Python / any HTTP client # Replace pm_live_xxxx with your actual key r = requests.get( "https://api.promptmatrix.io/pm/serve/your.prompt.key", headers={"Authorization": "Bearer pm_live_xxxx"}, timeout=2 ) system_prompt = r.text # returns the approved content

# After your LLM call — report the outcome back to PromptMatrix import requests, time # 1. Fetch prompt (you already do this) t_start = time.time() prompt = requests.get( "https://api.promptmatrix.io/pm/serve/agent.radar.system", headers={"Authorization": "Bearer YOUR_KEY"} ).text # 2. Run your LLM call (your existing code — unchanged) response = your_llm.call(prompt, user_input) latency_ms = int((time.time() - t_start) * 1000) # 3. Report outcome — ONE extra call, no LLM key needed requests.post( "https://api.promptmatrix.io/api/v1/agent/evals", headers={"Authorization": "Bearer YOUR_KEY"}, json={ "prompt_key": "agent.radar.system", # which prompt "score": 8.4, # your eval score (or omit for auto) "latency_ms": latency_ms, # how long the LLM took "tokens_in": response.usage.input, # input tokens "tokens_out": response.usage.output, # output tokens "model": "claude-sonnet-4-6", # model used "success": True, # did it succeed? "source": "agent", # "agent" | "product" | "test" # optional: "rationale": "response was coherent, on-topic" # optional: "session_id": "run-abc123" — group related calls } ) # That's it. PromptMatrix logs it, scores it, shows it in Performance tab. # If score < threshold → auto-flags for your Monday approval queue.