Agent Rules and Prompts

Use the current TeaQL Agent Kit for the full workflow. The short prompt below is intended to stay small enough for limited agent context windows. The generated application AGENTS.md and live, model-aware Assist provide the task-specific details.

Work model-first: evaluate and repair KSML before generation.
Read the generated application AGENTS.md.
Use entity/action Assist for the task at hand; request field-level Assist only
for fields you actually use. Never guess generated names.
Do not read or search generated domain-library source for API discovery.
If Assist is insufficient, report MISSING_ASSIST with the exact missing operation
and compiler diagnostic. Do not self-authorize source fallback.
Keep generated library files unchanged; implement in application-owned code.
Bound every query and include non-empty comment and purpose; purpose is immediately before execution.
Use an audit reason for every save; execute with trusted context.
Compile, test, run, and report exact commands, versions, failures, and skips.

API discovery in two steps

First ask for the entity's operation, then ask for a particular field if the operation requires it. For example:

cargo teaql rust-assist-query/school --input models/
cargo teaql rust-assist-query/school.established_date --input models/

Do not retrieve every entity or every Assist category up front. If the generated application already compiles and the task does not require a new business method, do not request unrelated create/update/delete Assist.

When something fails

Read the tool's Markdown diagnostic and the compiler error. A missing method calls for focused Assist, not a guessed spelling or a whole-file rewrite. Missing comment/purpose or audit intent is an application call-site problem; a template defect is reported upstream and verified by regeneration. Generated source may be inspected for framework debugging only under an explicitly approved, bounded source-fallback request—not as normal API discovery.

For acceptance, record the model version, generated target, selected runtime version, Assist calls, build/test/run exit results and any remaining MISSING_ASSIST. A green-looking console log without exit-zero evidence is not a completed verification.