Skip to content

armgit / macro

Public Imported from GitHub
0

Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.

2.4 KB
Raw Blame

Extraction prompts

All editable classification and extraction instructions live in this folder. Keep classification questions within Jev's 500-character limit; the intent prompt contract test enforces it. Describe intent boundaries in classification questions without copying behavior baseline inputs into those questions; use independently authored phrases to validate the rules rather than teaching the classifier each expected answer.

Files Used for
classify_<intent>_prompt.txt Independent intent questions; opening action words strongly signal the requested destination.
extract_search_unfiltered_query_prompt.txt Query-only search types retain unsupported filter clauses verbatim.
extract_fields_prompt.txt Shared instructions for generative content extraction.
extract_<intent>_prompt.txt Intent-specific generative instructions. Search adds selected-field guidance and the applicable query prompt.
typed_<intent>_<field>_prompt.txt Individual Clairic field instructions. Search uses the same per-type query prompts as generative extraction.
typed_search_scope_prompt.txt First-stage search type selection only.
search_<type>_<field>_prompt.txt Type-specific search filter guidance; field names and schemas derive from the selected Rust response model.
temporal_prompt.txt Temporal facts, source-span rules, and structured examples.
temporal_task_prompt.txt Task-specific deadline instructions appended to the temporal prompt.
temporal_thread_prompt.txt Deadline and closure facts read from an email thread for Focus; never used by Command K.
temporal_output_prompt.txt Output instruction preceding the generated temporal and thread fact schemas.

Rust assembles these instructions with generated schemas and task-property data. Prompt files are embedded at compile time; rebuild the harness after editing them. See the behavior baseline to compare live predictions.

Search type is selected first (or supplied by a manual Type choice), then only that type's response schema is extracted. Add new search fields to the typed Rust response model and supply field guidance here; the guidance coverage test fails when a model field has no prompt. Task search status/priority choices come from the frontend search picker catalog in property_definitions, not a prompt list.