Instructions
How to generate, review, edit, and choose tagging instructions.
Instructions
Instructions define how messages are tagged. A good instruction keeps categories stable, explains edge cases, and gives the model enough examples to handle real source text.
Main creation path
Use Auto-generation as the default way to create instructions. Upload already tagged examples and let the generator build the taxonomy, descriptions, and guidelines from real data.
After generation, use editing only for small corrections or advanced maintenance of an existing instruction.
Auto-generation
Use the generator when you have a CSV/XLSX file with tagged examples.
- Upload the examples file.
- Select brand, source system, and mapping mode.
- Verify detected mapping and extracted tag columns.
- Choose a policy mode:
- Precision for stricter tagging.
- Balanced for the default tradeoff.
- Coverage for broader recall. - Preview and confirm generation.
Review and editing
Open an instruction from the table to check its status, tags, and text. Edit names, tag descriptions, and guidelines when the taxonomy changes or model behavior needs adjustment.
Choosing an instruction for tagging
Pick an instruction that matches the brand, language, and source type. Reusing a mismatched instruction is one of the fastest ways to create noisy analytics.
Practical checks
- Tag names are short and consistent.
- Descriptions explain the boundary between similar tags.
- Examples include real ambiguous cases.
- The instruction is active before launch.