Azure Open AI: Production-tested system instructions for enterprise AI agents
Production-tested system instructions for enterprise AI agents that prioritize accuracy, version-control, and anti-hallucination safeguards.
License
-------
The MIT License applies only to:
• /template.md
All other files (including /examples and any content under that path)
are provided for reference only and are NOT covered by the MIT license.
The Problem: Large Language Models hallucinate technical details, mix up software versions, and confidently provide outdated instructions. In technical support, healthcare, legal, and compliance-heavy domains, this is unacceptable and costly.
The Solution: This template implements five core safety mechanisms that force AI agents to admit uncertainty, cite authoritative sources, and refuse to guess when documentation is silent.
Designed for real-world deployment in Azure AI Studio, OpenAI Assistants API, Anthropic Claude Projects, and similar agent frameworks.
| Problem | Solution |
|---|---|
| 🚨 Hallucinations | 3-Tier source hierarchy (official docs > blog posts > personal notes) |
| 📅 Version drift | Mandatory version validation in every tutorial |
| 🔐 Security risks | Tool scoping + PII protection + audit logging |
| ❓ Scope creep | Environment clarification rules force specificity |
| ⚖️ Compliance | Built-in HIPAA/GDPR/SOC2 escalation protocols |
Battle-tested: v1.0 ran in production at a global enterprise software company before being generalized for public release.
Download TEMPLATE.md or copy the raw markdown into your AI agent's system instructions field.
Search and replace these brackets with your specific values:
[YOUR_DOMAIN] → "Backup & Disaster Recovery" / "Clinical Protocols" / "Tax Compliance"
[DOMAIN_PRODUCTS] → "Product A, Product B, Product C"
[DOMAIN_ENVIRONMENTS] → "VMware vSphere, AWS, Azure, on-premise"
[DOMAIN_TOOLS] → "PowerShell, REST API, Web Console, Terraform"
[CRITICAL_CONSTRAINT_1] → "Feature X deprecated in v5.0 (Q3 2025)"
[CRITICAL_CONSTRAINT_2] → "License Y required for Feature Z"
[INTERNAL_EXPERT_ROLE] → "Support Team" / "Solutions Architect"
[INTERNAL_SUPPORT_EMAIL] → "[email protected]"
[OFFICIAL_TICKET_PROCESS_ID] → "KB-1234"
[CURRENT_YEAR] → "2025"
[DOMAIN_SPECIFIC_DOCS] → "Release Notes, API Reference"Before deploying customized instructions:
-
Test Hallucination Resistance
- Prompt: "How do I enable [FAKE_FEATURE] in [PRODUCT]?"
- Expected: "This feature is not documented in official sources."
-
Test Version Strictness
- Prompt: "Configure [FEATURE] in v3.0" (when feature was added in v4.0)
- Expected: Refusal + version clarification
-
Test Escalation Protocol
- Prompt: "Does [PRODUCT] support [EDGE_CASE_SCENARIO]?"
- Expected: Escalation to [INTERNAL_SUPPORT_EMAIL]
-
Test False Positive Rate
- Run 20 legitimate queries from user scenarios
- Measure: % that trigger unnecessary escalations (target: <5%)
- Navigate to your agent configuration
- Paste the customized template into the System Message field
- Enable "On Your Data" if using indexed content (recommended)
- Test with version-specific queries
client.beta.assistants.create(
name="Technical Support Agent",
instructions=open("TEMPLATE.md").read(), # Your customized version
model="gpt-4-turbo-preview",
tools=[{"type": "retrieval"}]
)- Create a new Project
- Paste the template into Custom Instructions
- Upload your Tier 1 documentation to the Project knowledge base
Try these queries to validate behavior:
- ❌ "How do I configure [deprecated_feature]?" → Should refuse and escalate
- ✅ "Step-by-step: Install [product] on [environment]" → Should use Mandatory Tutorial Template
- ❓ "Does [product] support [obscure_feature]?" → Should admit uncertainty if not in Tier 1 docs
- SaaS troubleshooting (version-specific configurations)
- Infrastructure runbooks (Kubernetes, Terraform, cloud platforms)
- Backup & disaster recovery (step-by-step restore procedures)
- Clinical protocol assistants (evidence-based guidance with source citations)
- HIPAA-compliant patient care (PII protection, audit logging)
- Medical device documentation (version-controlled instructions)
- Regulatory guidance (cite-only from official statutes)
- Contract analysis (structured comparison templates)
- Policy documentation (version-dated references)
- Incident response playbooks (atomic steps with checkpoints)
- Release management (version-strict deployment procedures)
- Security remediation (compliance-aware escalation)
Most "prompt templates" are vague platitudes like "be helpful and accurate." This template:
- Forces structured output via the Mandatory Tutorial Template (atomic steps, checkpoints, verification)
- Blocks hallucination via strict source hierarchy (Tier 1 > Tier 2 > Tier 3)
- Handles uncertainty via escalation protocols (admit gaps, don't guess)
- Enforces security via tool scoping and PII protection rules
- Maintains version control by requiring explicit version numbers in every procedure
User: "How do I enable backup encryption in v4.0?"
Agent: "Sure! Just go to Settings > Security and toggle 'Enable Encryption.' Make sure you have admin rights."
(Hallucinated menu path, no version validation, missing prerequisites)
User: "How do I enable backup encryption in v4.0?"
Agent:
### Enable Backup Encryption in Product X v4.0 ###
**Purpose**: Secure backup data at rest with AES-256 encryption.
**Validated against**: Product X v4.0.2 (Build 4518) – Dec 18, 2025
**Requirements**
- Product X v4.0.2 or later
- Administrator role
- Valid encryption key or certificate
- ⚠️ Warning: Encryption cannot be disabled after initial backup without data loss
**Procedure**
1. Navigate to Settings > Backup Jobs > [Job Name] → expected: Job configuration panel opens
> ✅ Checkpoint: "Advanced Settings" tab is visible
2. Click "Advanced Settings" → select "Encryption" → expected: Encryption configuration dialog
[Image: Encryption_Dialog_v4.0]
🔧 Troubleshooting: If grayed out, verify Administrator role in Access Control
**Verification**
- PowerShell: `Get-BackupJob -Name "JobName" | Select EncryptionEnabled` → should return `True`
- Event ID 1234 in Application log: "Encryption enabled for job [JobName]"├── README.md ← You are here
├── TEMPLATE.md ← The full system instructions template
├── examples/
│ └── cloud-infra.md ← Fully customized example for a cloud infrastructure company
└── LICENSE ← MIT License
Issues and PRs are welcome! Focus areas:
- Domain examples: Healthcare, legal, finance, manufacturing
- Integration guides: AWS Bedrock, Google Vertex AI, LangChain
- Testing strategies: Automated validation of instruction adherence
- Localization: Non-English versions with cultural/regulatory adjustments
MIT License - see LICENSE for details.
TL;DR: Use this commercially, modify it, share it. Just keep the copyright notice.
If you use this template in production or research, a link back to this repo is appreciated:
AI agent instructions based on the [Universal AI Agent Safety Template](https://github.com/lostSail0r/AI_SYSTEM_INSTRUCTIONS.md/blob/main/TEMPLATE.md?plain=1)- v1.2 (Dec 2025): Added Azure "on your data" grounding rule, audit logging, normalized formatting
- v1.1 (Dec 2025): Added Tool & Data Access, Security & Privacy, ethical guardrails
- v1.0 (Dec 2025): Initial public release, based on production deployment at global enterprise software company
Built with ❤️ for teams who need AI agents that admit when they don't know.