You're migrating a production prompt from Claude Opus 4.7 to GPT-5.5 to cut cost. What's the most reliable indicator that you can or cannot move the prompt as is?

Portability across models is an empirical question, not a theoretical one. The only reliable indicator is to run the prompt unchanged on both models against the same labeled set and compare. Build a golden set of 100 to 300 representative queries with known-good outputs from the current model (Claude Opus 4.7 in this case). Run the same prompt on GPT-5.5 and score every output with the metrics that matter for your task: faithfulness against retrieved context, format compliance against your output schema, refusal correctness on out of scope queries, end-task accuracy. Compare per-metric retention. A pass-gate of 95% faithfulness retention, 95% format compliance, and zero new false-refusals on the adversarial slice is a defensible 'it ports'. A meaningful drop on any axis means the prompt was over-tuned to Claude's quirks (XML tags, multi-paragraph system prompts, specific refusal calibration) and needs adaptation for GPT-5.5. The alternatives are worse. Eyeballing a few outputs is the classic vibes-driven loop that misses silent regressions on the cases you did not check. Matching context window and tokenizer is a necessary compatibility check, not a sufficient one; identical mechanics do not guarantee identical behavior. Reading both providers' prompt guides gives heuristics but not evidence on your task.

The only honest portability signal is an eval-driven comparison on the actual production task. **Setup.** Take a representative golden set (100 to 300 queries with known-good Claude Opus 4.7 outputs). Hold out a smaller adversarial slice (20 to 50 known-tricky queries that the current prompt handles well). Pick the metrics that matter for your application: faithfulness against retrieved context, format compliance against your schema, refusal correctness on out of scope, end-task accuracy, latency, and cost. **Measurement.** Run the unchanged prompt against both models. Score every output on every metric. Compute per-metric retention (GPT-5.5 score divided by Claude Opus 4.7 score on the same case) plus bootstrap confidence intervals so you can tell signal from noise. **Decision rule.** A typical defensible threshold is 95%+ faithfulness retention, 95%+ format compliance, zero new false refusals on the adversarial slice, and cost or latency at or below the current model's. If retention holds, ship the migration unchanged. If it drops meaningfully on any axis, the prompt was over-tuned to Claude's quirks and needs adaptation. **The adaptation patterns are predictable.** Claude prefers XML-tagged inputs (`<context>...</context>`) where GPT-5.5 tends to do better with markdown or JSON. Claude follows multi-paragraph system prompts reliably; GPT-5.5 sometimes needs the same content in a numbered list to lock in compliance. Refusal calibration differs: Claude tends to refuse more conservatively on borderline queries, GPT-5.5 less so, which shows up in your refusal-correctness metric. Knowing the adaptation patterns shortens the iteration cycle once the eval surfaces a gap. **The alternatives that fail.** Eyeballing five outputs misses silent regressions. Matching window and tokenizer is necessary but not sufficient. Provider docs give heuristics, not evidence. The eval is the only ground truth.