Best Practices
Prompt engineering patterns, document handling strategies,and data safety protocols for IAN Think.
Techniques for improving output quality, managing documents effectively, and operating within safety and compliance boundaries.
1. Prompt Construction
Prompt quality is the single largest variable affecting output quality. The patterns below are ordered by impact.
1.1 Specificity
Constrain the task. Specify the subject, scope, output length, and audience in every prompt.
| Low-specificity prompt | High-specificity prompt | What changed |
|---|---|---|
| "Tell me about claims" | "Walk me through how a homeowner files a water damage claim" | Added subject (homeowner), action (files), and scope (water damage) |
| "What's in this document?" | "Extract the financial projections for Q1–Q4 from this report" | Named the data type (financial projections) and range (Q1–Q4) |
| "Help me write something" | "Draft a 3-paragraph email to a client explaining a coverage denial. Tone: empathetic but firm." | Defined format (3 paragraphs), audience (client), topic (coverage denial), and tone |
1.2 Context Injection
IAN operates with zero background knowledge about the user's role, project, or organizational context. Provide this explicitly.
| Without context | With context |
|---|---|
| "Is this clause standard?" | "I'm reviewing a vendor contract for our claims department. Is the indemnification clause in Section 5 standard for our industry?" |
| "Rewrite this" | "Rewrite this for an audience of new hires who don't know industry jargon yet" |
Pattern — role declaration: Prepend a single sentence identifying your role and objective.
"I'm a claims adjuster training new team members. Explain subrogation in a way I can use in a training session."
"I'm in HR preparing for open enrollment. Summarize the key changes in our benefits plan for an all-hands presentation."
1.3 Output Format Specification
IAN mirrors the format described in the prompt. Explicitly request the structure you need:
"Give me a bulleted list""Format this as a table with columns: Risk, Impact, Likelihood""Write this as a numbered step-by-step procedure""Keep it under 100 words"
If no format is specified, IAN defaults to unstructured prose.
2. Iterative Prompting
2.1 Follow-Up Questions
IAN generates 3–5 suggested follow-up questions after each response. These are clickable shortcuts that continue the conversation in context.
Each message in a thread is additive — IAN retains the full conversation history within the session.
Example: building layered understanding
| Turn | Prompt |
|---|---|
| 1 | "What's a CAT bond?" |
| 2 | "Give me a real example" |
| 3 | "What are the risks for investors?" |
| 4 | "How does this compare to traditional reinsurance?" |
Each turn builds on prior context. This produces deeper, more coherent answers than isolated single-turn queries.
2.2 Correction and Redirection
If output quality is off-target, issue an explicit correction rather than starting over:
"That's too technical. Simplify it.""I need this from the policyholder's perspective, not the insurer's.""That's not quite right. The deadline is 30 days, not 60.""More detail on the second point. Less on everything else."
IAN adjusts within the existing thread. A new chat is not required.
2.3 When to Reset
Start a new chat when:
- The conversation exceeds ~20 messages without converging on a useful output.
- The topic has shifted substantially from the original prompt.
- IAN begins repeating or contradicting prior responses.
A clean session eliminates accumulated context noise.
3. Document Handling
3.1 Supported Formats
| Format | Suitability | Notes |
|---|---|---|
| PDF (text-based) | High | Selectable text required |
| DOCX | High | Native text extraction |
| CSV / XLSX | High | Optimal for structured and tabular data |
| Scanned image PDF | Not supported | IAN processes extracted text only. Run OCR before uploading. |
3.2 Source Verification
When IAN makes a claim about document contents, request a direct quote:
"Quote the exact text where you found that."
If IAN produces an accurate verbatim quote, confidence in the answer is higher. If it cannot cite a source passage, treat the output with increased skepticism.
3.3 Structural Hints
For documents with complex layouts (multi-column, nested tables, numbered sections), describe the structure explicitly:
"This is a table with columns: Date, Region, Revenue, Expenses. The first row is headers."
"Sections 1–3 cover current policies. Sections 4–6 cover proposed changes."
Structural hints reduce parsing errors and improve extraction accuracy.
3.4 Large Document Strategy
For documents exceeding ~50 pages, use a segmented analysis workflow:
- Split the document into 3–5 sections by topic.
- Upload one section per chat session.
- Summarize each section independently.
- Open a final chat and paste all summaries with the prompt:
"Synthesize these findings into one overview."
This avoids context window saturation and improves per-section accuracy.
4. Data Safety and Compliance
4.1 Output Verification
IAN can hallucinate — generate plausible but factually incorrect text. This is an inherent limitation of large language models, not a bug specific to IAN.
Mandatory verification domains: Always verify IAN output against primary sources before acting when the output involves:
| Domain | Examples |
|---|---|
| Financial | Budgets, pricing, settlement amounts |
| Regulatory / Legal | Compliance requirements, legal interpretations |
| Claims | Coverage determinations, liability assessments |
| Personnel | HR decisions, performance evaluations |
4.2 Data Classification
Before uploading any document, confirm its classification:
| Classification | Upload permitted? |
|---|---|
| Public | Yes |
| Internal | Yes |
| Confidential | Requires admin approval |
| Restricted | Requires admin approval |
4.3 PII Redaction
Redact personally identifiable information (names, Social Security numbers, account numbers) from documents before uploading when feasible.
4.4 Audit Visibility
All chat sessions are organization-owned data. Organization admins have read access to chat logs for compliance and audit purposes. Do not enter any content you would not want recorded in an audit log.
5. Session Management
5.1 Starring
Click ⭐ to pin a conversation. Starred chats are retained indefinitely and do not expire.
5.2 Exporting
Access via ⋮ (overflow menu) → Export.
| Format | Use case |
|---|---|
| Sharing via email, attaching to case files | |
| Markdown | Editing, embedding in reports or wikis |
5.3 Renaming
Access via ⋮ → Rename. Replace default titles (e.g., "New Chat") with descriptive names (e.g., "Q2 Budget Analysis", "Claims Procedure Questions") to support future retrieval.
Quick Reference — Five Core Habits
| # | Habit | Rationale |
|---|---|---|
| 1 | Be specific | Precision in the prompt directly correlates with precision in the output |
| 2 | Iterate | First-turn answers are rarely final. Follow-up prompts refine quality |
| 3 | Verify | Treat IAN output as a draft, not a source of truth |
| 4 | Save | Star chats for retention; export for distribution |
| 5 | Experiment | Vary models, phrasings, and formats to discover what works for your use case |