FetchXML Builder – AI Instructions

This folder contains the AI instruction files used by FetchXML Builder in XrmToolBox.

These files define how the AI assistant helps users build, modify, and explain FetchXML queries.
They are split into separate files to keep responsibilities clear and avoid prompt conflicts.

---

Instruction Set

These files are combined into the main instruction sent to the AI.

• System.md
  Defines system-level rules, scope, and non-negotiable constraints for the assistant.

• Behavior.md
  Describes how the AI uses metadata, updates FetchXML, handles execution, and explains queries.

• Style.md
  Controls tone, personality, and how the user is addressed.

• Preferences.md
  Contains naming conventions, output formatting rules, and interaction preferences.

• UserFlavors.md
  Injects optional user-specific guidance so the assistant can adapt to different user flavors.

Strictness profiles

One strictness file can also be included in the instruction set to control how cautiously or boldly the AI should interpret the user’s request.

• Strictness.Relaxed.md Favors progress and likely interpretations when the intent seems clear enough.

• Strictness.Balanced.md Balances helpful inference with brief verification when uncertainty matters.

• Strictness.Exact.md Favors literal interpretation and verification over guesswork.

---

Runtime context

At runtime, FetchXML Builder injects dynamic context into the instructions, such as:

• The current FetchXML query
• The user’s preferred name
• Schema metadata information
• The active strictness mode, when used

Placeholders (for example or) are populated by the tool and must not be modified.

---

Helper Prompt Files

These prompt files are used during sampling to help the AI resolve specific runtime subtasks.

• Updated.md Re-states the latest FetchXML and marks it as the authoritative current query for subsequent steps.

• EntityMeta.md Helps the AI match one requested table name or description to candidate Dataverse tables using supplied metadata only.

• AttributeMeta.md Helps the AI match one requested column description to candidate Dataverse columns for a known table using supplied metadata only.

• RelationshipMeta.md Helps the AI match one requested relationship description to candidate related tables and relationships for a known table using supplied metadata only.

---

Notes

• The AI does not have access to live Dataverse data
• FetchXML Builder sends schema metadata and current query context, not Dataverse record data
• All schema knowledge comes from schema metadata provided by FetchXML Builder
• Instructions are written in Markdown for readability and maintainability

---

Instruction architecture

The instruction files are treated as compiled templates, not dynamic prompts.

Before being sent to the AI model, FetchXML Builder:

1. Loads each instruction file
2. Replaces placeholders (for example ,, ``)
3. Sends the fully resolved instructions to the model

As a result:

• The AI never sees placeholders
• All preferences are already decided
• Instructions are interpreted as factual and authoritative

This design avoids conditional logic in prompts and produces more deterministic behavior.

---

Links

This page: https://rappen.github.io/Tools/AI/FXB/

FetchXML Builder: https://fetchxmlbuilder.com/

FXB AI Chat: https://fetchxmlbuilder.com/features/ai/

---

FXB AI Log Example

This example will show both the initial instructions, and a few prompts between the user and the AO assistant.

FXB AI Log Example.md