Coding Prompts
Prompt Migration for Current Models
Paste an old prompt written for GPT-4o, Claude 3.x, or Gemini 1.5/2.0 and get back a version rewritten for current reasoning models.
Prompt
Rewrite this prompt for a current reasoning model (it was originally written for an older, non-reasoning model). Current models like GPT-6, Claude Fable 5.1/Opus 5.5/Sonnet 5, and Gemini 3.x think internally before answering, so they need different handling than the model this was written for.
OLD PROMPT:
[PASTE THE OLD PROMPT]
ORIGINALLY WRITTEN FOR: [e.g. "GPT-4o" or "Claude 3.5 Sonnet"]
TARGET MODEL: [e.g. "GPT-6 Sol" or "Claude Sonnet 5"]
When rewriting:
1. Remove any "think step by step," "let's work through this carefully," or scripted reasoning instructions ("first do X, then verify Y, then compute Z") — the model already reasons internally, and this scaffolding can override its own better plan.
2. Remove any `<thinking>` tag instructions asking the model to show its reasoning in the visible output — that duplicates the hidden reasoning phase and costs tokens twice.
3. Keep or strengthen: the task description, constraints, output format, and any "verify your answer" instruction (verification instructions still help).
4. Flag anything in the old prompt that depended on model-specific quirks that may no longer apply (e.g. a workaround for a limitation the new model doesn't have).
5. Suggest an appropriate reasoning effort level (low/medium/high/xhigh) for the task, and note the parameter name for the target model's API.
Return the rewritten prompt, then a short bullet list of what changed and why.
How to use
Run this once per prompt in your prompt library when migrating off an older model. It's meant for prompts you know behaviorally worked but might now be fighting the new model rather than working with it — not for prompts you're writing fresh, which should just follow current best practice from the start.
Variables
[PASTE THE OLD PROMPT]— the exact prompt text, including any system prompt portion[ORIGINALLY WRITTEN FOR]— helps the model calibrate how much scaffolding to expect and strip[TARGET MODEL]— the specific model and tier you're migrating to, so the effort-level suggestion is concrete
Tips
- Don't trust the rewrite blindly on a high-stakes prompt. Run both versions against a small test set (Promptfoo makes this a one-config-file job) and compare before replacing the old one in production.
- If the old prompt was doing a lot of format-teaching through worked examples (few-shot with reasoning shown), that few-shot pattern for format usually survives migration fine. This prompt is mainly about stripping reasoning-scaffolding, not gutting every example.
- For a large prompt library, batch this: run it across every prompt file, but review the diffs yourself before merging. A model can occasionally strip an instruction that looked like scripted reasoning but was actually a genuine business rule.