Prompt Optimizer

Congratulations on the promotion. The "generic" feeling almost always comes from missing context, not missing creativity. Your prompts likely describe what to write, but not who is writing, who is reading, or what makes your product different.

First, inject a defined brand persona. Instead of "write a product description," specify: "You are a witty, minimalist outdoor-gear copywriter who speaks in short sentences and never uses exclamation points." Give the AI your brand's tone, vocabulary, and forbidden words.

Second, add audience specificity. "Appealing to busy moms" is weak. Say: "Speak to a 34-year-old urban parent who values durability over style, has 10 minutes to shop, and hates jargon." This forces concrete choices.

Third, supply proof points and differentiators. Generic copy happens when the AI invents benefits. List your actual specs, materials, and customer reviews. Tell it: "Use these three unique selling points, and only these."

Finally, demand structure and constraints. Ask for a hook, a pain-point paragraph, a feature-benefit list, and a call-to-action with your specific discount code. Constraints breed specificity. Add a self-check: "Read back and flag any sentence that could apply to a competitor." That single line transforms output quality.

Structure for Protocol-Aligned Categorization

Your issue is common: generic summarization prompts treat feedback as flat text, while your protocol demands a severity taxonomy. Fix this by encoding your ranking system directly into the prompt, not hoping the model infers it.

Define the Severity Scale Explicitly

List your exact levels (e.g., Grade 1–4, or Mild/Moderate/Severe) with operational definitions. Avoid vague words like "serious." Instead, write: "Grade 2 = requires medical intervention but no hospitalization." The model needs your rubric, not its own assumptions.

Force Structured Output

End your prompt with a strict format requirement, such as: "Return a markdown table with columns: Feedback ID, Category, Severity Grade, Justification (one sentence quoting the source text)." This prevents the model from blending or omitting severity.

Add a Verification Step

Include a final instruction: "Before outputting, check each entry against the severity definitions above. If uncertain, assign the lower grade and flag it." This reduces over- and under-rating.

Test with Anchored Examples

Provide 2–3 sample feedback items with your correct labels. Few-shot examples dramatically improve consistency. Ask the model to follow the same reasoning pattern.

Would you like me to draft a full template incorporating your specific grade definitions?

The variance you're seeing almost always comes from ambiguity in the prompt, not the model's mood. To get consistent labeling and sorting, you need to decompose the task into discrete, verifiable steps that leave zero room for interpretation.

Start by defining your role explicitly. For example: "You are a data-cleaning specialist for survey responses." Then, separate the workflow into numbered stages: 1) Detect and flag missing values, 2) Standardize categorical labels (e.g., 'M'/'F' to 'Male'/'Female'), 3) Sort by a specified column in ascending order. Each step should have a rigid output format, like "Return a table with columns: ID, Original_Response, Cleaned_Label, Sort_Order."

The key is to avoid vague verbs like "clean" or "sort properly." Instead, specify the exact rules for each transformation. If you want deterministic behavior, also constrain the output structure—e.g., "Always output a markdown table, never prose." Finally, add a self-check step: "Before returning, verify that all labels match the allowed set and that sorting is applied to the cleaned column only." This forces the model to follow a linear path rather than improvising.

If you're still seeing drift, test with a fixed seed or temperature setting, but prompt structure is your primary lever.