Prompt builder
Most bad support prompts fail the same few ways. Build one from parts and the checker flags those failures as you type. Try the "Bad prompt" template first.
0/ 100
How it works
- No model is involved. Every check is a plain rule in prompt-builder.js, so you can see exactly why something is flagged.
- Customer text sits in its own tagged block, before the instructions. The model can tell your instructions from the customer's words, and long text sits where models handle it best.
- The biggest weights go to the task, the output format and the "missing information" rule. Those three are the easiest to get wrong and the most expensive when you do.
- Token count is a rough estimate (characters รท 4). The Tokens page gives exact counts.
In production I would
- Keep prompts in version control with an owner, like code. A prompt change is a release.
- Run every change against a fixed set of real, anonymised tickets before it ships, and compare scores.
- Add a model-based review on top of these rules for things rules miss, such as tone and policy fit.
- Log which prompt version produced each answer, so a bad reply can be traced back.