When to give up on a prompt

Some prompts cannot be fixed by iteration. After three or four rounds of editing, if the output is still wrong in the same direction, the underlying approach is wrong — and the discipline is to abandon the prompt and start over with a different structure, not to keep editing.

Signs the prompt is structurally wrong

When to switch patterns

If a Pattern 1 (respond-in-voice) prompt is not working well for guest emails, try Pattern 2 (extract-from-text) first to extract the guest's actual request, then a separate Pattern 1 to respond. The two-step pattern usually outperforms the single-step pattern for complex inputs.

If a Pattern 4 (generate-with-rules) prompt is producing too much variation, switch to Pattern 1 with strict voice constraints — generation tasks often benefit from being framed as response tasks where the "response" is the generated content.

When to switch models

Some tasks just work better on specific models. Claude tends to be better at long-context tasks (review responses with 200-message history). GPT-4 tends to be better at structured extraction (JSON output). Gemini tends to be better at multimodal tasks (image + text). If your prompt is well-structured but the output is still wrong, switching from one model to another sometimes solves the problem instantly.

When to give up entirely

Some hospitality use cases genuinely do not work with current LLMs. Examples: "calculate the optimal rate for tomorrow given these 14 segment forecasts" (LLMs are bad at this — use an RMS), "predict whether this guest will cancel" (use a trained predictive model, not an LLM), "translate a complex legal contract without errors" (LLMs are good but not good enough for legal — use a sworn translator). Recognizing the use cases where AI is not the right tool saves the team from spending months on a workflow that will never reach production quality.