Constrain classifiers to JSON and 'other'

Use a request that asks for intermediate work you can inspect: classify only the supplied comment into billing, setup, reliability, or other, return schema-valid JSON, and route uncertainty to other. A useful prompt says what the tool may use, the required format, and what it should do when evidence is missing. It does not demand secret internal reasoning or reward confident guessing.

Try a two-pass loop. First ask for the structured artifact and questions. Then correct one concrete problem using the original evidence. This preserves the distinction between the source and the instruction. If the result contains a claim you cannot trace, ask it to mark the claim unsupported rather than rewrite it more smoothly.

Write a short record before you move on. For classifying customer feedback into a human-reviewed weekly summary that is never auto-sent, state the claim or choice under review, then name the evidence that supports it: the original comment, a versioned label guide, JSON validation, a reviewer decision, and weekly aggregate counts. Next, name what the evidence does not establish. This last line prevents a narrow test from becoming a broad promise. If another team member opened your record next month, they should be able to reproduce the review without trusting your memory.

Now make the trade-off visible. A support lead running a low-risk workflow with a small operations team may value speed, clarity, cost, control, or reassurance differently. Explain which of those mattered in the current version and why. Do not let a model choose the trade-off simply because it can produce a confident answer. The responsible owner decides whether the upside justifies the remaining uncertainty.

Finally, connect the decision to a next action. If the current evidence is enough, identify the smallest safe step forward. If it is not, request a specific source, approval, or test. The stop condition remains concrete: the output is not valid JSON, confidence is low, the comment contains sensitive data, or any step would email, refund, or update a customer. A documented pause is a successful outcome when it keeps a weak result from becoming a consequential one.