> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxis-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Best Practices

> A Comprehensive Guide to the Top 10 Best Practices

This guide shows how to get the most out of Pria with precise prompts, complete context via IP Vault, verifiable outputs with Tool Details, and repeatable workflows using Assistants. It aligns with Praxis AI's operating model and links to relevant documentation where applicable.

To unlock the full potential of Praxis AI's Pria and ensure optimal, precise, and efficient interactions, integrate these ten best practices into your workflow. By doing so, you can significantly enhance the quality of responses, streamline research, and effectively manage your AI-driven tasks.

***

## At a glance

<CardGroup>
  <Card title="1- Be specific" icon="bullseye" href="#1-be-specific%3A-the-power-of-detailed-requests-and-context">
    Craft precise, bounded prompts with clear deliverables, tone, and constraints.
  </Card>

  <Card title="2- Use IP Vault for websites" icon="globe" href="#2-website-research-via-ip-vault%3A-comprehensive-scraping">
    Scrape full sites to give Pria clean, complete context instead of noisy partial HTML.
  </Card>

  <Card title="3- Verify with Tool Details" icon="magnifying-glass" href="#3-tool-details-unveil-process-and-verification">
    See which tool ran, sources used, success status, and returned data.
  </Card>

  <Card title="4- Manage conversations" icon="comments" href="#4-manage-conversations%3A-organize-context-and-history">
    Create and choose the right conversation to preserve context and retrieve history.
  </Card>

  <Card title="5- Set perspective" icon="user-gear" href="#5-use-perspective%3A-define-prias-role-explicitly">
    Tell Pria which role to adopt for audience-aligned output.
  </Card>

  <Card title="6- Create Assistants" icon="robot" href="#6-create-assistants%3A-automate-and-optimize-workflows">
    Automate multi-step workflows with clear, ordered instructions and error handling.
  </Card>

  <Card title="7- Apply CRISPE" icon="wand-magic-sparkles" href="#7-error-handling-patterns-for-assistants">
    Design robust Assistants with a structured capability and context template.
  </Card>

  <Card title="8- Iterate & validate" icon="check-double" href="#8-advanced-design-with-crispe">
    Use Tool Details to identify gaps and refine prompts or sources accordingly.
  </Card>

  <Card title="9- Stay organized" icon="folder-open" href="#9-iterate-and-validate-with-a-tight-loop">
    Name conversations and Assistants consistently for easy retrieval.
  </Card>

  <Card title="10- Respect governance" icon="shield" href="#10-governance%2C-privacy%2C-and-safety">
    Follow data, privacy, and safety rules using IP Vault and RAG over ad-hoc copy-paste.
  </Card>
</CardGroup>

***

## 1 Be Specific: The power of detailed requests and context

<Info>
  The foundation of effective interaction with Pria lies in the specificity and depth of your prompts. A well-crafted request, rich in detail and relevant context, directly correlates with the quality and accuracy of the AI's response. Specificity directly improves precision and reduces rework.
</Info>

Think of your prompt as a detailed instruction manual for Pria. The more precise you are, the better the results.

### Why Specificity Matters

Avoid vague or ambiguous language. Clearly articulate what you need, including any specific parameters, desired formats, or limitations. A detailed request helps Pria understand exactly what you're looking for and deliver accurate, relevant responses.

### Building Specific Prompts

<Steps>
  <Step title="State the objective and audience">
    What should be produced and for whom (e.g., teachers vs. executives). Define the purpose and target audience clearly.
  </Step>

  <Step title="Constrain format and scope">
    Word count, headings, components, style, and what to exclude. Set boundaries for length, structure, and content.
  </Step>

  <Step title="Provide source context">
    Attach RAG/IP Vault content or quote key excerpts. Give Pria access to the information sources it should use.
  </Step>

  <Step title="Define success criteria">
    Checks, tests, or acceptance conditions. Explain how to evaluate if the output meets your needs.
  </Step>
</Steps>

### Example: Generic vs. Specific

<Tip>
  **Generic:** "Write about marketing"

  **Specific:** "Write a 500-word blog post on the latest trends in digital marketing for small businesses, focusing on SEO strategies and social media engagement, in a conversational yet authoritative tone."
</Tip>

### Complete Example Prompt

```txt theme={null}
Objective: Draft a 500–600 word announcement for small businesses about Q1 SEO trends.
Audience & tone: Conversational yet authoritative; non-technical.
Format: 4 sections (Overview, Trend 1, Trend 2, Actions); include 3 actionable bullets.
Constraints: Cite any stats inline; avoid jargon like "synergy"; no images.
Sources: Use my IP Vault document for "AcmeSEO Blog Q1 2025" and my RAG file "Search Trends Notes".
Success: Clear 90-day action plan and linked glossary terms.
```

***

## 2 Website research via IP Vault: Comprehensive scraping

When gathering information from websites, leverage the IP Vault for comprehensive data acquisition rather than simply asking Pria to "look at this website." This method offers significant advantages in terms of content depth and cleanliness.

### Why IP Vault?

<CardGroup>
  <Card title="Full-context capture" icon="download">
    Pulls complete site content for deep analysis, not just partial HTML. Utilizing the IP Vault to scrape a site allows you to capture the entire content, providing Pria with the complete context needed for thorough analysis and accurate response generation.
  </Card>

  <Card title="Clean data" icon="wand-magic-sparkles">
    Filters out extraneous JS/CSS and delivers structured text. Direct URL reading by Pria typically yields only about 20% of the content, often intermingled with extraneous HTML and JavaScript. Scraping via the IP Vault ensures a much cleaner, more comprehensive dataset.
  </Card>

  <Card title="Repeatable" icon="rotate">
    Documents are reusable across conversations and Assistants. Once scraped, the content can be referenced multiple times without re-scraping.
  </Card>
</CardGroup>

### The Problem with Direct URL Reading

Direct URL reading by Pria often results in:

* Only \~20% of actual content captured
* Content mixed with HTML, JavaScript, and CSS code
* Incomplete context for analysis
* Noisy, unstructured data

### Using IP Vault Effectively

<Steps>
  <Step title="Create a site snapshot file">
    Add the site to IP Vault and complete the scrape. This captures the full website content in a clean, structured format.
  </Step>

  <Step title="Name and tag consistently">
    Use descriptive names like `VendorX Docs (2025-10-02) | pricing, api, auth` to make documents easy to find and reference.
  </Step>

  <Step title="Reference documents explicitly">
    Tell Pria which file to use by name to ensure it uses the right source.
  </Step>

  <Step title="Limit scope as needed">
    Ask for specific sections to avoid noise and focus on relevant content.
  </Step>

  <Step title="Refresh on updates">
    Re-scrape periodically when content changes to keep your data current.
  </Step>
</Steps>

### Example Usage

```txt theme={null}
Use IP Vault: "Compare pricing tiers from 'VendorX Docs (2025-10-02)' and map features to our needs.
Avoid direct URL reads; use my document only."
```

<Tip>
  This eliminates the limitations often encountered when Pria directly reads URLs and improves Pria's ability to process and synthesize information effectively.
</Tip>

***

## 3 Tool Details: Unveil process and verification

<Info>
  The "Tool Details" feature is an invaluable tool for understanding how Pria constructs its responses and verifying the integrity of the information provided. Tool Details shows which tool executed, whether it succeeded, which sources were used (RAG, IP Vault, browser), and what data was returned.
</Info>

This transparency is crucial for assessing the reliability and originality of the AI's output.

### What Tool Details Reveals

<AccordionGroup>
  <Accordion title="Understanding Response Compilation" icon="puzzle-piece">
    By examining Tool Details, you can discern which sources and methods Pria employed to compile its response. This insight helps you differentiate between information directly retrieved (e.g., from a search in RAG via `call_rag` or the internet via `get_browser`) and content that might be inferred or generated by the LLM.

    **Key information includes:**

    * Tool identity and version
    * Execution duration
    * Input parameters used
    * Server label (mcp)
    * Response Text
  </Accordion>

  <Accordion title="Assessing Success and Returns" icon="clipboard-check">
    The details indicate which tool was used, whether it was successful in its operation, and what specific data or output it returned. This allows you to:

    * Verify successful execution
    * Identify error messages and failure points
    * Review source breakdown (RAG sections, IP Vault documents, web fetches)
    * Examine returned artifacts (tables, JSON, files)
  </Accordion>

  <Accordion title="Source Verification" icon="magnifying-glass">
    Tool Details shows exactly which sources were consulted:

    * RAG documents and specific sections
    * IP Vault files referenced
    * Web searches performed
    * API calls made

    This transparency helps you validate that Pria used the correct and most relevant sources.
  </Accordion>

  <Accordion title="How to act on gaps" icon="wrench">
    Use the information in Tool Details to improve your interactions:

    * **If sources are missing:** Attach or reference the right RAG/IP Vault items
    * **If a step failed:** Retry with narrower scope or corrected parameters
    * **If inference is high:** Ask Pria to separate "retrieved facts" vs "model inferences"
    * **If output quality is low:** Identify which steps succeeded and which need refinement
  </Accordion>
</AccordionGroup>

### Why This Matters

Tool Details helps you identify potential gaps in information retrieval or areas where Pria might have had to "fill in the blanks," guiding you on when further investigation or clarification might be needed.

<Note>
  Always review Tool Details for critical or high-stakes outputs to ensure the response is built on reliable, verifiable sources rather than model inference.
</Note>

***

## 4 Manage conversations: Organize context and history

Effective conversation management is essential for maintaining context and preventing information loss within your interactions with Pria. Each new thought or topic should ideally initiate a new conversation.

### Why Conversation Management Matters

<Info>
  Pria's responses are heavily influenced by the current conversation's history. Selecting the appropriate conversation ensures that Pria has access to all relevant prior interactions, leading to more coherent and accurate responses.
</Info>

### Contextual Integrity

Conversations organize your dialogues and maintain the context that Pria uses to generate responses. Proper conversation management ensures:

* **Context Preservation:** Pria remembers relevant details from earlier in the conversation
* **Accuracy:** Responses build on established context rather than starting fresh
* **Coherence:** Multi-turn interactions maintain logical flow and reference earlier points

### Best Practices for Conversation Management

<Steps>
  <Step title="Start a new conversation per topic">
    Prevents cross-topic contamination. Each distinct topic or project should have its own conversation thread.
  </Step>

  <Step title="Pick the correct conversation before asking">
    Ensures relevant history is available. Before asking a question, verify you're in the right conversation context.
  </Step>

  <Step title="Name conversations descriptively">
    Use clear, specific names like "Mintlify Writer — Pria Best Practices" or "Q4 Marketing Campaign Planning" to make conversations easy to identify and retrieve.
  </Step>

  <Step title="Leverage 'Clear selected' to broaden history">
    Allows Pria to draw across prior threads when appropriate. This broadens Pria's contextual understanding, allowing it to draw upon a wider range of past interactions.
  </Step>
</Steps>

### Organization and Retrieval

<Note>
  If something "disappears," check a different conversation thread. Context is scoped per thread by design. If you suspect information has been lost, it might simply be part of a different conversation thread.
</Note>

### Assistant-Driven Workflows

Conversations are often created in conjunction with Assistants, which are designed to execute specific instructional workflows. Aligning your questions with the correct conversation ensures Pria applies the appropriate Assistant's logic and context.

<Tip>
  Create a naming convention for your conversations that includes the project, date, or Assistant type to make them easier to find later.
</Tip>

***

## 5 Use perspective: Define Pria's role explicitly

Directing Pria to adopt a specific persona or role can significantly enhance the relevance and tone of its responses. By framing your request with a defined perspective, you guide Pria to generate output that aligns with that particular expertise.

### Why Perspective Matters

Explicitly stating Pria's desired role ensures the output resonates with the intended audience and purpose. The perspective you set determines:

* **Vocabulary and terminology** used
* **Depth and technical level** of explanations
* **Tone and communication style**
* **Focus and priorities** in the response

### Examples of Effective Perspective Setting

<CardGroup>
  <Card title="Social Media Expert" icon="hashtag">
    ```txt theme={null}
    "As a social media strategist, create a 7-day content plan prioritizing short-form video
    for small retailers. Include hooks, CTAs, and platform-specific notes."
    ```
  </Card>

  <Card title="AI Engineer" icon="code">
    ```txt theme={null}
    "As an AI engineer, outline a Node service to summarize PDFs using our IP Vault documents.
    Include architecture, error handling, and timeouts."
    ```
  </Card>

  <Card title="Business Analyst" icon="chart-line">
    ```txt theme={null}
    "As a business analyst, review our Q3 sales data and identify the top 3 trends
    impacting revenue. Present findings in executive summary format."
    ```
  </Card>

  <Card title="Technical Writer" icon="book">
    ```txt theme={null}
    "As a technical writer, document this API endpoint with clear examples,
    parameter descriptions, and common error scenarios."
    ```
  </Card>
</CardGroup>

### Tailored Communication

When you instruct Pria to act as an "AI engineer," it will leverage its understanding of AI principles and development to "build an app for me that does A, B, and C," providing solutions and insights relevant to that technical domain.

### Setting Effective Perspectives

<Tip>
  State domain, deliverable, constraints, and evaluation criteria in the first message to lock in the perspective. This ensures consistent output aligned with your needs.
</Tip>

**Template for perspective setting:**

```txt theme={null}
As a [ROLE], [ACTION] for [AUDIENCE/PURPOSE].
Include [SPECIFIC ELEMENTS].
Use [TONE/STYLE].
Focus on [PRIORITIES].
```

<Info>
  The perspective you set influences every aspect of Pria's response, from word choice to structural organization to the depth of technical detail provided.
</Info>

***

## 6 Create Assistants: Automate and optimize workflows

The ability to create custom Assistants within Praxis AI is a powerful tool for automating repetitive tasks and streamlining complex operational workflows. Developing this skill is a crucial aspect of mastering prompt engineering.

### Why Create Assistants?

<CardGroup>
  <Card title="Automation" icon="gears">
    Assistants are designed to follow a defined set of step-by-step instructions, effectively automating routine or multi-stage processes. This frees up your time for more strategic tasks.
  </Card>

  <Card title="Consistency" icon="check-circle">
    Structured operational workflows ensure consistency and accuracy in task execution across multiple runs.
  </Card>

  <Card title="Repeatability" icon="arrows-rotate">
    Turn reliable processes into repeatable, multi-step runs with consistent quality.
  </Card>

  <Card title="Optimization" icon="gauge-high">
    Refine and optimize workflows over time based on performance and outcomes.
  </Card>
</CardGroup>

### Assistant Core Pattern

Assistants turn reliable processes into repeatable, multi-step runs with consistent quality. An Assistant's instructions define a clear, sequential workflow for the AI to follow.

### Basic Assistant Template

```txt theme={null}
You are the "XXX" assistant and your role is to create YYY by following the step-by-step instructions below carefully:

1- Collect input from user through conversation
2- Analyze the data
3- Generate your response
4- End of job

Error Handling:
A- Not enough or invalid input data: Continue or restart at step 1
B- Off topic: Redirect off topic back to the focus of this assistant
```

### Key Elements of Effective Assistants

<Steps>
  <Step title="Clear role definition">
    Start by defining what the Assistant is and what it produces.
  </Step>

  <Step title="Sequential steps">
    Break down the workflow into numbered, atomic steps that are easy to follow and verify.
  </Step>

  <Step title="Error handling">
    Define what happens when things go wrong or inputs are invalid.
  </Step>

  <Step title="Termination condition">
    Specify when the Assistant's job is complete.
  </Step>
</Steps>

<Info>
  Keep steps atomic, verifiable, and source-aware (RAG/IP Vault names). Add logging expectations and acceptance checks for robust operation.
</Info>

### Advanced Considerations

When designing Assistants:

* **Name sources explicitly:** Reference specific RAG or IP Vault documents by name
* **Include validation steps:** Check inputs before processing
* **Define success criteria:** What does "done correctly" look like?
* **Plan for edge cases:** How should the Assistant handle unexpected inputs?

<Tip>
  Start with simple Assistants and gradually add complexity as you understand the patterns that work well for your use cases.
</Tip>

***

## 7 Error handling patterns for Assistants

Building robust Assistants requires thoughtful error handling and validation strategies. This ensures your Assistants can gracefully handle unexpected situations and provide useful feedback.

### Core Error Handling Patterns

<CardGroup>
  <Card title="Validate inputs" icon="circle-check">
    Check required fields and data types before processing. Ask targeted follow-ups when information is missing or unclear.
  </Card>

  <Card title="Constrain sources" icon="filter">
    Prefer specific RAG/IP Vault documents over free web searches. This ensures consistent, reliable information sources.
  </Card>

  <Card title="Plan-then-execute" icon="clipboard-list">
    Have the Assistant outline its plan before executing. This allows for validation and course correction early in the process.
  </Card>

  <Card title="Degrade gracefully" icon="triangle-exclamation">
    Return partial results plus a "next steps" list when complete execution isn't possible. Never fail silently.
  </Card>
</CardGroup>

### Error Categories to Handle

<AccordionGroup>
  <Accordion title="Input Errors" icon="keyboard">
    **Scenario:** User provides incomplete, invalid, or ambiguous input.

    **Response Strategy:**

    * List specific fields that are missing
    * Provide examples of valid inputs
    * Ask targeted questions to gather needed information
    * Offer to restart from step 1 with clear inputs
  </Accordion>

  <Accordion title="Source Errors" icon="file-circle-xmark">
    **Scenario:** Referenced documents or data sources are not available.

    **Response Strategy:**

    * Identify which sources are missing
    * Suggest alternative sources if available
    * Ask user to provide or specify the correct sources
    * Explain what can and cannot be completed without the sources
  </Accordion>

  <Accordion title="Processing Errors" icon="gear">
    **Scenario:** An intermediate step fails or produces unexpected results.

    **Response Strategy:**

    * Identify which step failed and why
    * Present what was successfully completed
    * Offer options to retry with different parameters
    * Suggest breaking the task into smaller steps
  </Accordion>

  <Accordion title="Scope Errors" icon="expand">
    **Scenario:** User request is off-topic or outside the Assistant's capabilities.

    **Response Strategy:**

    * Acknowledge the request
    * Explain the Assistant's specific focus and capabilities
    * Redirect to the intended workflow
    * Suggest alternative Assistants or approaches if applicable
  </Accordion>
</AccordionGroup>

### Example Error Handling Implementation

```txt theme={null}
Error Handling Protocol:

A- Invalid or Incomplete Input:
   - Identify missing required fields
   - Provide examples of valid input format
   - Ask: "Please provide [specific field] in [format]. Example: [example]"
   - Offer to restart from step 1

B- Source Not Available:
   - Report: "Cannot access [source name]"
   - Ask: "Please verify the document name or provide an alternative source"
   - List available alternatives if any exist

C- Processing Failure:
   - Report what was completed successfully
   - Explain where the failure occurred
   - Suggest: "We can try [alternative approach] or adjust [parameters]"

D- Off-Topic Request:
   - Acknowledge: "I understand you're asking about [topic]"
   - Clarify: "This Assistant focuses specifically on [scope]"
   - Redirect: "For [off-topic], please use [appropriate resource/Assistant]"
```

<Note>
  Well-designed error handling transforms frustrating failures into learning opportunities and improves the overall user experience.
</Note>

***

## 8 Advanced design with CRISPE

For more sophisticated and robust Assistants, explore the CRISPE methodology. This advanced template provides a structured approach to defining your Assistant's capabilities, context, and desired outputs.

### Why CRISPE?

<Info>
  CRISPE is a structured methodology to define Capabilities, Role, Inputs, Steps, Policies, and Evaluation for robust, auditable Assistants. It ensures comprehensive Assistant design that covers all aspects of operation.
</Info>

### The CRISPE Framework

<Steps>
  <Step title="Capabilities & constraints">
    What the Assistant can and cannot do. Define the boundaries of functionality clearly to set appropriate expectations.
  </Step>

  <Step title="Role & audience">
    Who it is and who it serves. Establish the persona and target user base for appropriate tone and depth.
  </Step>

  <Step title="Inputs & sources">
    Required fields and named RAG/IP Vault assets. Specify exactly what information the Assistant needs to operate effectively.
  </Step>

  <Step title="Steps & checkpoints">
    Atomic tasks with intermediate validations. Break down the workflow into verifiable stages with clear success criteria.
  </Step>

  <Step title="Policies & safety">
    Privacy, redaction, and compliance boundaries. Establish guardrails for data handling and output generation.
  </Step>

  <Step title="Evaluation & QA">
    Acceptance tests and quality thresholds. Define how to measure success and what constitutes acceptable output.
  </Step>
</Steps>

### CRISPE Template Structure

```txt theme={null}
Assistant: [NAME]

CAPABILITIES:
- Can: [list specific capabilities]
- Cannot: [list explicit limitations]
- Requires: [dependencies and prerequisites]

ROLE:
- Identity: [who/what the Assistant is]
- Serves: [target audience]
- Purpose: [primary objective]

INPUTS:
- Required: [mandatory fields and formats]
- Optional: [nice-to-have inputs]
- Sources: [specific RAG/IP Vault documents by name]

STEPS:
1. [Step name]: [detailed action]
   - Checkpoint: [verification criteria]
2. [Step name]: [detailed action]
   - Checkpoint: [verification criteria]
[Continue for all steps]

POLICIES:
- Privacy: [data handling rules]
- Safety: [content boundaries]
- Compliance: [regulatory requirements]
- Errors: [handling strategies]

EVALUATION:
- Success criteria: [measurable outcomes]
- Quality metrics: [standards to meet]
- Acceptance tests: [validation steps]
```

### Benefits of CRISPE Design

<CardGroup>
  <Card title="Comprehensive" icon="list-check">
    Covers all aspects of Assistant behavior and operation systematically.
  </Card>

  <Card title="Auditable" icon="clipboard">
    Creates clear documentation of what the Assistant should do and how.
  </Card>

  <Card title="Maintainable" icon="screwdriver-wrench">
    Makes it easy to update and refine Assistant behavior over time.
  </Card>

  <Card title="Reliable" icon="shield">
    Reduces edge cases and unexpected behaviors through thorough design.
  </Card>
</CardGroup>

<Note>
  Learn more about building Assistants with CRISPE: [Building Assistants with CRISPE](/mdx/user-guide/assistants-personalization/building)
</Note>

<Tip>
  Start with the basic template and evolve to CRISPE as your Assistants become more complex or mission-critical.
</Tip>

***

## 9 Iterate and validate with a tight loop

Effective use of Pria requires an iterative approach where you continuously refine your prompts and validate outputs using Tool Details and other feedback mechanisms.

### The Iteration Loop

<Steps>
  <Step title="Initial request">
    Start with a clear, specific prompt based on best practices.
  </Step>

  <Step title="Review output">
    Evaluate the response for accuracy, completeness, and alignment with your needs.
  </Step>

  <Step title="Check Tool Details">
    Examine which sources were used, which tools executed, and whether all steps succeeded.
  </Step>

  <Step title="Identify gaps">
    Determine what's missing, incorrect, or could be improved.
  </Step>

  <Step title="Refine and retry">
    Adjust your prompt, add sources, or clarify requirements based on findings.
  </Step>

  <Step title="Repeat until satisfied">
    Continue the loop until outputs meet your acceptance criteria.
  </Step>
</Steps>

### Validation Request Template

```txt theme={null}
"Run with sources A/B. In Tool Details, show: (1) retrieved passages,
(2) any inference steps, (3) unresolved items. If missing, request the exact inputs needed."
```

### What to Look For

<AccordionGroup>
  <Accordion title="Source validation" icon="check">
    * Were the right documents consulted?
    * Are there missing sources that should be included?
    * Is the source information current and accurate?
  </Accordion>

  <Accordion title="Completeness" icon="list">
    * Are all requested elements present?
    * Is the depth appropriate for your needs?
    * Are there gaps in coverage or reasoning?
  </Accordion>

  <Accordion title="Accuracy" icon="bullseye">
    * Are facts verifiable and correctly cited?
    * Is there a clear distinction between retrieved facts and inferences?
    * Do numbers, dates, and specific details check out?
  </Accordion>

  <Accordion title="Quality" icon="star">
    * Does the tone match your requirements?
    * Is the structure logical and easy to follow?
    * Does it meet your stated success criteria?
  </Accordion>
</AccordionGroup>

### Refinement Strategies

When iteration reveals issues:

<CardGroup>
  <Card title="Add specificity" icon="crosshairs">
    Make your requirements more precise and bounded.
  </Card>

  <Card title="Adjust sources" icon="book">
    Add, remove, or replace reference documents.
  </Card>

  <Card title="Break down tasks" icon="scissors">
    Split complex requests into smaller, sequential steps.
  </Card>

  <Card title="Clarify constraints" icon="ruler">
    Make explicit what should and shouldn't be included.
  </Card>
</CardGroup>

<Info>
  Treat outputs as drafts until Tool Details confirms the right sources were used and success criteria are met. This iterative validation ensures high-quality, reliable results.
</Info>

<Tip>
  Keep a log of what refinements worked well for different types of tasks. This builds your personal library of effective prompt patterns.
</Tip>

***

## 10 Governance, privacy, and safety

Operating Pria responsibly requires attention to data governance, privacy protection, and safety considerations. These practices protect both you and your organization while ensuring compliant, ethical AI use.

### Core Governance Principles

<AccordionGroup>
  <Accordion title="Prefer structured sources" icon="database">
    Use IP Vault documents and RAG files instead of raw copy-paste or unvetted links. Structured sources provide:

    * Better traceability and audit trails
    * Controlled access and versioning
    * Cleaner, more reliable data
    * Easier compliance with data policies
  </Accordion>

  <Accordion title="Respect data boundaries" icon="shield-halved">
    Follow your organization's privacy and safety policies:

    * Restrict sensitive content appropriately
    * Verify recipients for any exports
    * Use proper classification labels
    * Follow data retention policies
    * Respect geographic and regulatory boundaries
  </Accordion>

  <Accordion title="Human-in-the-loop" icon="user-shield">
    Maintain human review for outputs impacting:

    * Critical business decisions
    * Grades or evaluations
    * Compliance or legal matters
    * Policy creation or changes
    * Public communications
    * Financial transactions

    AI assistance should augment, not replace, human judgment in high-stakes scenarios.
  </Accordion>

  <Accordion title="Transparency and disclosure" icon="eye">
    Be transparent about AI usage:

    * Disclose when content is AI-generated
    * Cite sources appropriately
    * Explain AI's role in decision-making
    * Document processes for auditability
  </Accordion>
</AccordionGroup>

### Data Privacy Best Practices

<Steps>
  <Step title="Classify before sharing">
    Understand the sensitivity level of data before using it with Pria.
  </Step>

  <Step title="Use IP Vault for control">
    Store sensitive documents in IP Vault where access can be managed and tracked.
  </Step>

  <Step title="Minimize PII exposure">
    Redact or generalize personally identifiable information when possible.
  </Step>

  <Step title="Review before distribution">
    Check outputs for inadvertent disclosure of sensitive information.
  </Step>
</Steps>

### Safety Considerations

<Info>
  Safety encompasses both preventing harmful outputs and ensuring appropriate use of AI capabilities.
</Info>

**Content safety:**

* Review outputs for bias, fairness, and appropriateness
* Ensure compliance with content policies
* Validate factual claims, especially for sensitive topics
* Consider potential misuse or misinterpretation

**Operational safety:**

* Test Assistants thoroughly before production use
* Implement fallbacks for critical workflows
* Monitor for unexpected behaviors or errors
* Have contingency plans for AI unavailability

### Compliance Framework

Different industries and regions have specific requirements:

* **Education:** FERPA, COPPA, student privacy laws
* **Healthcare:** HIPAA, patient confidentiality
* **Finance:** SOX, PCI-DSS, financial regulations
* **Europe:** GDPR and data protection rules
* **Government:** FedRAMP, ITAR, specific agency requirements

<Note>
  For institutional commitments and protections, see your Praxis AI Data Processing Agreement (DPA) and internal governance policies.
</Note>

### Ethical AI Use

Beyond compliance, consider ethical dimensions:

* **Fairness:** Avoid perpetuating biases or discrimination
* **Accountability:** Take responsibility for AI-assisted outputs
* **Transparency:** Be clear about AI's role and limitations
* **Beneficence:** Use AI to create positive outcomes
* **Autonomy:** Respect user agency and informed consent

<Tip>
  When in doubt about governance, privacy, or safety, consult your organization's AI governance team or compliance officer before proceeding.
</Tip>

***

## Putting it all together

Mastering Pria requires integrating all these practices into a cohesive workflow. Here's how to combine them effectively:

### The Complete Workflow

<Frame>
  <Steps>
    <Step>
      Start a correctly named conversation
    </Step>

    <Step>
      Set a clear perspective
    </Step>

    <Step>
      Select precise files from IP Vault
    </Step>

    <Step>
      State objective, constraints, and success criteria
    </Step>

    <Step>
      Run
    </Step>

    <Step>
      Inspect Tool Details
    </Step>

    <Step>
      Refine inputs or steps
    </Step>

    <Step>
      Repeat until acceptance tests pass
    </Step>

    <Step>
      Templatize into an Assistant with CRISPE
    </Step>

    <Step>
      Add fallback for edge cases
    </Step>
  </Steps>
</Frame>

### Continuous Improvement

As you use Pria more:

1. **Build a prompt library:** Save successful prompts and patterns
2. **Refine Assistants:** Update based on real-world performance
3. **Document learnings:** Note what works and what doesn't
4. **Share knowledge:** Help others learn from your experience
5. **Stay current:** Follow Praxis AI updates and new capabilities

### Troubleshooting Common Issues

<CardGroup>
  <Card title="Generic outputs" icon="triangle-exclamation">
    **Solution:** Raise specificity (objective, structure, constraints) and narrow sources (IP Vault/RAG).
  </Card>

  <Card title="Incomplete results" icon="puzzle-piece">
    **Solution:** Check Tool Details for missing or failing steps and adjust accordingly.
  </Card>

  <Card title="Wrong context" icon="map-marked">
    **Solution:** Verify you're in the correct conversation and the right sources are referenced.
  </Card>

  <Card title="Inconsistent quality" icon="chart-line">
    **Solution:** Create an Assistant to standardize the workflow and ensure repeatability.
  </Card>
</CardGroup>

<Tip>
  If results feel generic, raise specificity (objective, structure, constraints) and narrow sources (IP Vault/RAG). If results feel incomplete, check Tool Details for missing or failing steps and adjust accordingly.
</Tip>

### Success Indicators

You're mastering Pria when:

* First attempts consistently produce usable outputs
* You can predict which sources and steps Pria will use
* Assistants handle 80%+ of routine workflows
* Tool Details shows successful execution with minimal inference
* Iteration cycles are shorter and more targeted
* Outputs require less manual editing
* Colleagues ask you for tips and templates

***

## Related documentation

Explore these resources to deepen your Praxis AI expertise:

* **Getting Started:** [User Guide](/mdx/introduction/welcome) - Foundation for using Praxis AI
* **Building Assistants:** [Assistants Personalization](/mdx/user-guide/assistants-personalization/building) - Advanced Assistant design with CRISPE
* **IP Vault:** [Managing Your IP Vault](/mdx/user-guide/files-documents/ipvault) - Learn how to effectively scrape and manage website content with Pria's RAG
* **Prompt Engineering:** [Assistants Overview](/mdx/user-guide/assistants-personalization/assistants) - Master the art of crafting effective prompts and building custom assistants
* **Tools:** [Tools Reference](/mdx/admin-guide/tools) - Understand the capabilities of the AI tools available in Pria

<Tip>
  Bookmark this guide and refer back to it regularly as you develop your Praxis AI skills. Mastery comes through consistent application of these practices.
</Tip>