Even with rules, custom GPTs forget to follow instructions. This keeps them on track.
Standardize instructions for better quality outputs.
Many communicators are comfortable using custom GPTs to streamline workflows and create better outputs. They’re developing extensive prompt libraries for improved results.
Sometimes, though, custom GPTs forget instructions or start hallucinating nonsense, said Sarah Evans, partner and head of PR at Zen Media. She’ll be speaking more about this next month at Ragan’s PR Daily Conference.
Evans pointed to the widely debated em dash as an example.
“Even though I’ve put it (in my instructions) a million times…the draft almost always has em dashes,” she said.
The issue isn’t necessarily your instructions or the technology. It’s a lack of systems in place to achieve the best and most accurate results, she said.
“You can’t magically just say, ‘hey, ChatGPT, build me a GPT for optimizing press releases,’ because it has to come from your knowledge,” Evans said.
Teams getting the most meaningful and consistent results from AI are documenting how their organizations think, write, edit and review content, then feed those standards back into the system for every task, she said.
Standardize the rules
Before building a custom GPT or workflow, define the standards your teams are already using internally. This is everything from editorial guidelines, formatting rules, trusted sources and banned phrases, Evans said.
The goal should be to turn institutional knowledge into a repeatable infrastructure, she said.
“It has to come from your knowledge,” Evans said. “That’s where AI hasn’t taken over.”
This process has become especially important as communicators optimize content for AI search and LLMs. Evans said press releases and owned media now need different structures if companies want them surfaced in AI-generated answers.
“If (LLMs are) only going to read the first 240 characters, you might have to do unique things now,” she said, including adding multiple subheads or FAQ sections designed around audience prompts.
But Evans said many teams stop too early with their instructions. They build a GPT once, assume it understands their standards and move on.
Add an addendum every time
Instead, her team repeatedly reinforces instructions through attached documents and addendums for every single task.
One example is a detailed editorial PDF Evans attaches to prompts.
Learn more about this from Ragan’s Communications Leadership Council here.
Evans’ document includes pages of instructions, including:
- “AI slop” reduction rules
- Banned phrases and “AI tells”
- Hallucination guidelines
- Headline standards
- Source trust maps
- Lede formulas
- Compliance instructions
“I attach it to every piece of content,” Evans said. “Repetition matters for each task. Redundancy is everything.”
Restate instructions, attach reference files repeatedly and then break tasks into step-by-step workflows.
“It works better when you tell it first do this, second do that,” she said. “You kinda have to talk to it like it’s a toddler.”
The process may sound tedious, but Evans argues the payoff comes when those standards scale across your organization.
Putting it into practice
One example came from a junior writer on her team who spent roughly two hours researching articles before drafting them.
Evans mapped the employee’s workflow, questioned each step and turned the process into a reusable AI skill she attached as an addendum to the prompt. The team then tested the system repeatedly to identify errors and weak points.
“She’s now down to, call it 10 to 12 minutes, on that same skill now,” Evans said.
This workflow process is now shared across her editorial, social and PR teams, she said.
“It is a one- or two-time investment that is now saving our team literally hundreds of hours,” she said. “Everyone thinks they’re behind. I think I’m behind. We are building and learning at the speed of technology right now. The rules are gonna keep changing. We need to adjust with it.”
Courtney Blackann is a communications reporter. Connect with her on LinkedIn or email her at [email protected].
