Built-in instructions
The built-in instructions set the basic rules for the agent to work with Engee. They are transmitted along with descriptions of performing tools, and hints about launching, authorization, and skills are provided in the tool responses. Therefore, the user does not need to manually set a large system instruction before each task.
These are instructions for the agent, not a means of proving correctness. The result is still checked by stored artifacts, diagnostics and calculation data.
The difference from skills
| Layer | Content | When used |
|---|---|---|
Built-in instructions |
General rules for the API, saving models, retrieving data, and diagnosing |
When working with the appropriate tools |
Skills |
The methodology of a specific task, reference files, checks and the format of the result |
After selecting and downloading via |
User Task |
Requirements, restrictions, input data, acceptable actions |
During the current task |
Below are the basic rules of operation for the AI agent, established by the built-in instructions Engee MIND.
-
* First, the state of the environment.* The agent calls
engee_status; the start of the stopped environment is performed viaengee_startwhen a task requires calculations. -
Public API. For models used
engee.*with a lowercase letter.eval_codeaccepts Julia code, not Python or MATLAB. -
Creating a model up to blocks. First
engee.create, then adding blocks and links. Binary.engeeyou can’t replace it with a manually generated text file. -
Verified blocks and parameters. The block path is obtained via
block_search, parameters and ports — viablock_info. Values forengee.set_param!passed in strings. -
Regular saving. Creating a model in a session does not create a file automatically. The model is saved immediately after creation and after significant changes, for example
engee.save("MyModel", "/user/MyModel/MyModel.engee"; force=true). Overwriting is allowed only for the selected working file. -
A readable diagram. The agent sets the location of the blocks and checks the connections. Automatic layout should not degrade the feedback circuit.
-
Data of the current run. Blocks are used to receive signals via Engee MIND To Workspace with unique names and
collect(variable name_). The lack of data cannot be replaced by plausible numbers. -
Diagnostics in the source environment. If the model is not calculated, the agent reads the diagnostics, checks the parameters and corrects the model Engee. Replacing the model with an external calculation requires a separate user decision.
-
Explicit file transfer. The file in the session and the file on the client’s computer are different objects. To download and receive files, use
file_uploadandfile_download.
What happens when a calculation error occurs
The agent checks for the block To Workspace, uniqueness of variable names, initialization logs, connections, parameters and solver settings. An empty launch result does not prove that the circuit has worked correctly. For analysis, the following is used engee_diagnose, and for visual inspection — engee_screenshot an open model.
Numbers are checked for finiteness, units of measurement and physical plausibility. Large values are estimated relative to the scale of the task: by themselves they do not prove an error. Matching series when comparing different implementations require checking the signal sources.
What you need to ask the user
Specify the model or project, unit parameters, initial conditions, time range, acceptable changes, and acceptance criteria. For example:
В модели `/user/Drive/Drive.engee` найди причину отсутствия данных скорости. Сохрани исправленную копию в `/user/Drive/Drive_fixed.engee`. Физические параметры не меняй. Покажи причину ошибки, данные повторного прогона и сравнение с ожидаемым установившимся режимом.