Prompt iteration on a live-web tool is guesswork. You rewrite the instructions, the answer changes, and two things moved at once — your prompt and the pages the model happened to read that minute. Without a baseline you are tuning on noise.
The agent
| system_prompt | Summarise the document: what is new, what it replaces, the obligations it creates and who they bind, the transition period. Cite every claim. Say when the text is ambiguous instead of resolving it. |
| tool_config | web search, 10 fetches, no domain restriction |
| output_schema | free text with citations |
The workflow
Publish a baseline
v1 is your control. Run your standard question set against it and keep the runs.
Edit into a draft
Rewrite the prompt in the draft, test-drive it in a scratch chat, publish v2 with a changelog line.
Diff in replay mode
Rerun the same input on v2 against v1's stored sources. The web is frozen, so the only variable is your prompt.
Keep the verdict
Mark each comparison better / worse / equivalent. The roll-up across 40 inputs is your evidence, not a vibe.
What you get
Pin an agent. Run it in a year. Get the same thing back.
Claude, ChatGPT and Gemini included — no provider account to open. Build one agent, publish v1, run it, then change the prompt and diff the two.