Custom AI: Build a Risk Framework Assistant with Claude Projects

For Risk Analysts

Tools: Claude | Time to build: 1–2 hours | Difficulty: Intermediate-Advanced Prerequisites: Comfortable using Claude for document analysis — see Level 3 guide: "Analyze Regulatory Documents with Claude"

What This Builds

A persistent AI assistant that already knows your organization's risk taxonomy, appetite statements, policies, and control frameworks before you ask it anything. Instead of re-explaining your context in every conversation, you open the project and it already knows you're at a regional bank with a 45% CRE concentration and a specific risk appetite for operational losses. Every question you ask gets answers calibrated to your actual framework — not generic industry advice.

Prerequisites

• Comfortable using Claude for document analysis (Level 3)
• Claude Pro subscription ({{tool:Claude.price}}/month at {{tool:Claude.url}})
• Your organization's key risk documents ready: risk appetite statement, risk taxonomy/policy, KRI library, control framework, or any risk register
• Time to build: 1–2 hours initially; 15–20 minutes to add new documents later

The Concept

A Claude Project is like hiring a new analyst who reads your entire risk framework before their first day. Every conversation they have uses that background knowledge. You set it up once with your key documents and instructions — and then every chat in that project starts from shared context. Think of it as the difference between calling a random consultant vs. calling one who has already spent a week reviewing your policies.

Build It Step by Step

Part 1: Create the Project

1. Log in at {{tool:Claude.url}} and look for Projects in the left sidebar (may appear as a folder icon or "Projects" label)
2. Click New Project
3. Name it something clear: "Risk Framework Assistant — [Your Organization Name]"
4. You'll see two areas: Project Instructions (the system prompt) and Knowledge (documents you upload)

Part 2: Write Your Project Instructions

This is the most important step. Click Edit Project Instructions and write:

You are a Risk Framework Assistant for [Organization Name], a [organization type and size, e.g. $2B community bank in the Midwest].

Your role: Help our risk analysts work faster and more consistently using our actual risk framework.

Organization context:
- Key risk categories: [list your main risk types]
- Risk appetite: [brief summary, e.g. "Conservative; board-approved limits for credit concentration, operational loss, and liquidity"]
- Regulatory environment: [e.g. "OCC-regulated community bank; subject to FDIC deposit insurance; not subject to DFAST"]
- Key business lines: [list 2-3 key business lines]

What you help with:
1. Drafting risk report narrative sections (match our formal tone)
2. Generating KRI commentary from metric inputs
3. Creating first-draft risk register entries using our taxonomy
4. Reviewing vendor questionnaire responses
5. Summarizing regulatory documents and mapping to our existing policies
6. Drafting risk policy language consistent with our existing framework

Always:
- Use our risk taxonomy categories (not generic industry categories)
- Reference our documented risk appetite when assessing risk levels
- Flag when something might require legal or compliance review
- Be explicit when you're making an assumption vs. drawing from uploaded documents

Never:
- Make up regulatory citations
- Assert that specific regulatory requirements apply without flagging that legal review is needed

Click Save.

Part 3: Upload Your Knowledge Documents

Click Add to Project Knowledge and upload your key documents:

1. Risk Management Policy (the main framework document)
2. Risk Appetite Statement
3. KRI Library / Definition Sheet
4. Risk Taxonomy or Risk Register template
5. Any regulatory guidance most relevant to your work (OCC bulletins, Fed guidance)

For each document, Claude will show you it was uploaded. Supported formats: PDF, Word, plain text.

What you should see: Your project now shows the knowledge files listed. Each file icon shows it's been processed.

Part 4: Test and Refine

Start a new conversation within the project and test with a real task:

Here's our Q1 credit risk data: [paste 5-6 bullet points of metrics]. Write the credit risk section for our quarterly risk committee report. Match the tone and structure of our risk committee reports.

Review the response. Does it use your terminology? Does it reference the right risk categories? If not, refine the project instructions to be more specific.

Part 5: Add Context for Common Workflows

Add instructions for your most common tasks by editing the project instructions:

For risk report writing:
- Standard section length: 200-250 words
- Always start with the current period's headline metric, compare to prior period, then explain the key driver, then note control actions taken or planned
- Sign off with a forward-looking sentence

For KRI commentary:
- Format: [Metric name]: [Current value] vs. [Prior period] ([threshold]). [1-2 sentence explanation of trend and any actions.]

Real Example: Full Risk Committee Prep

Setup: Project contains risk policy, KRI definitions, and risk appetite statement

Input: "Here's our Q1 data for operational risk: [12 bullet points of loss event data, KRI values, control assessment results]"

Output: A structured risk committee narrative section — using the organization's specific terminology, organized by the right categories, with trend analysis matched to the uploaded KRI definitions

Time saved: 3 hours of report writing → 25 minutes of AI drafting + 20 minutes of review and editing

What to Do When It Breaks

• Claude doesn't use the uploaded documents: Check that files uploaded successfully (no error icons). Try asking explicitly: "Based on our KRI definitions document, how is [KRI name] defined?"
• Responses don't match your tone/format: Add more specific style instructions to the project instructions — paste an example paragraph from a prior report and say "Match this format"
• Claude makes up facts about your organization: It's inferring from the instructions, not the documents. Add the specific facts to the instructions explicitly
• Project knowledge is outdated: When policies change, delete the old document from project knowledge and upload the new version

Variations

• Simpler version: Use Claude's regular conversation with a long context-setting message (no persistent project) — works for one-off tasks but requires re-entering context each time
• Extended version: Add your loss event database or risk register as a CSV file to project knowledge — Claude can then reference actual historical events when generating reports or KRI commentary

What to Do Next

• This week: Build the project, upload 2–3 key documents, and test it on your next due report
• This month: Expand the knowledge base with more documents; refine instructions based on where responses feel off
• Advanced: Create a second project for TPRM work only — with vendor assessment templates, scoring frameworks, and common vendor questionnaire types as knowledge documents

Advanced guide for risk analyst professionals. Claude Projects require a paid subscription. All uploaded documents are processed by Anthropic — review your organization's data handling policies before uploading sensitive internal documents.