Best Practices

Custom GPTs built within the Ruder Finn ChatGPT enterprise workspace are intended for internal Ruder Finn use only. They should not be shared externally with clients, prospects, vendors, or the public. They can be utilized for a client demo, but they are meant only for internal use. Before you build or deploy a GPT, make sure it meets these baseline standards.

Access and Sharing

Set the access to “Only me” during building and testing, and only expand to your team once it’s ready. Never share a custom GPT outside of RF.

If you’re building something that you think could have broader internal value, please flag it to the team so it can be reviewed and potentially shared across our Enterprise account. This will help avoid duplication and enable us to benefit from one another’s creations.

Ownership & Accountability

Every GPT should have a clear owner. The person who built it is responsible for keeping it accurate and up to date. If you leave a team or project, hand off ownership to a member of your account team, or the AI Accelerator. Orphaned GPTs with stale information are worse than no GPT at all.

When to Use a Custom GPT (vs. Just Using ChatGPT)

A custom GPT makes sense when you find yourself giving ChatGPT the same context, background, or instructions over and over. If there’s a repeatable workflow, a consistent audience, or a body of internal knowledge that would make responses meaningfully better, that’s the signal to build one.

  • Answering recurring team questions using internal documents or frameworks
  • Drafting consistent communications in a defined format or voice
  • Summarizing or synthesizing a specific type of content on a regular basis
  • Testing pitches against reporter or client personas
  • Reformatting content into approved templates (press release, POV, briefing doc, etc.)

  • Translating technical information into plain-language explanations
  • Reviewing content for alignment with brand voice or messaging pillars

 

Skip the custom GPT if the task is one-off, highly variable, or doesn’t benefit from persistent context. Standard ChatGPT is the right tool for open-ended, exploratory, or ad hoc work.

Always build your GPT with a clearly defined audience in mind. GPTs that are too broad, or try to accomplish too much within one GPT are not as successful as those with clear targets and objectives.

Building Useful Instructions for a Custom GPT

“Instructions” are the behind-the-scenes manual telling a GPT how to operate. These instructions

  • Should define the GPT’s role and primary function
  • Detail how it should retrieve and apply knowledge
  • Outline behavior guidelines about what it should do and what it should avoid
  • Explain format and structure requirements
  • Provide voice and tone direction
  • Detail explicit do’s and don’ts.

 

Instructions serve as the memory for a GPT. When your instructions ask a GPT to refer to knowledge base documents, make sure to use clear, predictable labels for those documents (e.g., “Voice Rules,” “Good Examples”, “Formats”, etc.) so the GPT can retrieve them more reliably.

Building Useful Knowledge Base

The knowledge base is what makes a custom GPT genuinely valuable. It is what separates it from a generic AI response. Think of it as a reference guide. But quality matters more than quantity.

  • Upload documents that are current and authoritative. Outdated SOPs or old pitch decks will actively make outputs worse.
  • Use clean, readable files. PDFs with complex formatting, scanned documents, or heavily image-based files often don’t parse well. Plain text, Word docs, and well-structured PDFs work best.
  • Be selective. A GPT trained on 3 highly relevant documents will outperform one loaded with 30 loosely related ones. More is not better.
  • Keep it maintained. Set a reminder to review your knowledge base every quarter. When source documents are updated outside of the GPT, update them in the GPT knowledge base too.
  • Name your files clearly. The GPT uses filenames as context — “Q4_2025_Playbook_FINAL.docx” is more useful than “Document1.pdf.”

A Few Final Reminders

  • Responses from a GPT, or “outputs” are a starting point only, not a final product. Always review and provide your strategic thinking on an output before using or sending.
  • A GPT is a living product that must be refined and tested over time. Do regular quality checks, both self-reviews and reviews with colleagues, on the outputs.
  • The GPT is only as good as the instructions and knowledge you give it. If results are off, the instructions are usually where to look first.
  • If a GPT you’ve built gets meaningful adoption, document it – what it does, who it’s for, and what it shouldn’t be used for.
  • When in doubt, ask. Reach out to your team’s dedicated AI enthusiast or the AI Accelerator before building something you’re not sure about.
  • If you have a GPT you think others in the agency could benefit from, please send a note to the AI Accelerator team: aiaccelerator@ruderfinn.com.

 Questions or want to flag a GPT that should be shared more broadly? Contact the AI Accelerator aiaccelerator@ruderfinn.com.