Building a Lending Assistant: Custom GPTs & Claude Projects

So far you've used AI one prompt at a time. The next level is building a reusable assistant that already knows your products, your templates, and your tone — so you stop re-explaining context every time. Custom GPTs (in ChatGPT) and Projects (in Claude) let you package your best prompts and reference materials into a dedicated lending assistant you can use all day. This lesson shows you how, and where the compliance lines are.

What You'll Learn

• The difference between Custom GPTs and Claude Projects
• How to build a lending assistant step by step
• Three high-value assistants worth building first
• How to keep a reusable assistant private and compliant

Custom GPTs vs. Claude Projects

Both solve the same problem: giving an AI persistent context so it behaves like a specialist instead of a generalist.

• Custom GPTs (ChatGPT, paid plans) let you write standing instructions, upload reference files (your product matrix, checklist templates, a style guide), and reuse the assistant with one click. You can keep them private to yourself.
• Claude Projects (Claude, paid plans) work similarly: a project has its own knowledge files and custom instructions, and every conversation inside it starts with that context loaded. Claude's large context makes Projects especially good for document-heavy lending work.

For most loan officers, the choice comes down to which tool your institution has approved. The build process is nearly identical.

Building Your Lending Assistant: Step by Step

Step 1 — Define its job. Pick one clear purpose. "Borrower communications assistant" or "credit memo drafter" beats a vague "banking helper."

Step 2 — Write the standing instructions. This is the assistant's permanent system prompt. Example for a borrower-communications assistant:

You are a communications assistant for a consumer loan officer. Your
job is to draft clear, warm, compliant borrower emails and product
explanations.

Always:
- Use plain language a borrower with no finance background understands.
- Keep a warm, professional, reassuring tone.
- Label any example numbers clearly as illustrative, never as a quote.

Never:
- State a specific interest rate, APR, or payment amount.
- Make any promise about approval or future outcomes.
- Include any customer's personal identifying information.

If a request would require quoting a rate or making a promise, remind
me to pull the verified figure from our system instead.

Step 3 — Add knowledge files. Upload redacted, non-confidential reference materials: your product one-pagers, your email templates, your document checklists, your house style guide. Now the assistant answers in your context.

Step 4 — Test and refine. Run real (redacted) scenarios. If it drifts — gets too formal, forgets a rule — tighten the instructions. Treat the instruction block as living documentation.

Three Assistants Worth Building First

1. Borrower Communications Assistant. Knowledge: product sheets, email templates, FAQ answers. Use: drafts emails, explanations, and check-ins in your voice all day.

2. Credit Memo Drafter. Knowledge: your memo template, a strong sample memo (redacted), your section order and tone. Use: paste redacted underwriting notes, get a committee-ready draft in house style. Build in the "use only my facts, invent no numbers" rule from the standing instructions.

3. Document & Checklist Helper. Knowledge: your per-product checklists and stipulation templates. Use: generates tailored checklists and precise stipulation requests, and runs completeness reviews against the right list.

Each of these encodes the lessons from earlier modules so you're not rebuilding context every time.

Sharing — Carefully

You can share Custom GPTs and Projects with colleagues, which is tempting for standardizing a team's communications. Two cautions:

• Only share through your institution's approved, data-protected AI environment. A personal consumer account is not the place for shared bank workflows.
• Knowledge files must be non-confidential. Product sheets and blank templates are fine; never load real customer files, populated applications, or anything with NPI into a shared assistant.

The Compliance Reality of Reusable Assistants

A reusable assistant is more powerful — which means its mistakes are more repeatable. A bad instruction gets applied to every borrower email until someone catches it. So:

• Bake the guardrails into the standing instructions (no rates, no promises, no NPI, label examples) so they apply automatically.
• Review output the same way — a Custom GPT draft is still a draft you verify and own.
• Keep a human in every consequential loop — declines, credit decisions, and disclosures never become "the assistant's job."
• Update it as products and rules change. An assistant trained on last year's product sheet will confidently use last year's terms.

Built well, a lending assistant is like a well-trained junior teammate who never forgets your templates and never gets tired — but who still needs you to check the work and sign the file.

Key Takeaways

• Custom GPTs (ChatGPT) and Claude Projects give an AI persistent context — your products, templates, and tone — so it acts like a specialist.
• Build one with a clear job, strong standing instructions (with guardrails baked in), and non-confidential knowledge files; then test and refine.
• Start with a Borrower Communications Assistant, a Credit Memo Drafter, and a Document/Checklist Helper.
• A reusable assistant repeats its mistakes too — bake in guardrails, review every output, keep a human in consequential loops, and update it as products and rules change.