A mid-sized agency had rolled out an AI assistant to draft first-pass replies to client support tickets. On paper it was a win: response drafts appeared in seconds, and agents only had to review and send. In practice, the support lead kept hearing the same complaint from clients — the replies felt cold, scripted, and faintly corporate. Agents were rewriting nearly every draft before sending, which erased the time savings the tool was supposed to deliver.
This is a story about fixing that problem with role prompting alone — no new model, no fine-tuning, no extra tooling. It is also a story about how the obvious fix ("make it sound friendlier") was not the fix at all, and how a structured persona redesign got the team to a draft agents could actually send. The names are generic, but the arc is one we see constantly: a capable model producing the wrong voice because nobody told it who to be.
What follows is the situation, the decision, the execution, the outcome, and the lessons that transfer to any team in the same spot.
The Situation
The assistant's original prompt was about as bare as prompts get.
The Starting Prompt
"Draft a reply to this support ticket: [ticket]."
No persona, no tone guidance, no constraints. The model produced grammatically perfect, accurate, and utterly generic replies. They opened with "Thank you for reaching out" and closed with "Please don't hesitate to contact us," every single time. Accurate, but lifeless.
The Measurable Problem
The support lead tracked one number: the share of AI drafts sent without edits. It sat near 8 percent. Agents were rewriting more than nine out of ten drafts, which meant the tool was adding a review step without removing the writing step.
The Decision
The team's first instinct was to add "be friendly and warm" to the prompt. They tried it. It made things worse — now the replies were generically cheerful, peppered with exclamation marks that clashed with the brand's calm, competent tone.
Reframing the Goal
The breakthrough was reframing from "friendlier" to "who would write the ideal reply." They described an actual person: a support specialist who had been with the company for years, knew the product cold, and wrote like a calm, capable human. That shift, from adjective to persona, is the heart of our best practices that actually work.
The Execution
They rebuilt the prompt using the four-layer structure — identity, context, priorities, constraints.
The Redesigned Persona
"You are a senior support specialist who has worked at this company for six years. You write like a calm, competent human: warm but never gushing, clear, and concise. Acknowledge the specific issue in the first sentence, give the concrete next step, and avoid generic phrases like 'reaching out' and 'don't hesitate.' No exclamation marks. Keep it under 120 words."
Why Each Piece Mattered
- The tenure detail conditioned a confident, knowledgeable register.
- The banned phrases killed the robotic openers and closers directly.
- The "specific issue first" rule forced the reply to engage with the actual ticket instead of opening with boilerplate.
They placed the persona in the system message so it held across every ticket, following the placement logic in our step-by-step approach to role prompting.
The Outcome
The team measured the same metric over the following weeks.
The Numbers Moved
The share of drafts sent without edits rose from roughly 8 percent to around 55 percent. Agents still reviewed every draft, but most now needed only a light touch instead of a full rewrite. The tone complaints from clients faded.
What Did Not Change
Accuracy stayed flat — as expected. The persona never touched correctness; it changed voice and structure. When the model lacked information to answer a ticket, it still lacked it. The team handled that separately by giving agents a clear "escalate" path, a reminder that personas fix tone, not knowledge, echoing our 7 common mistakes with role prompting.
The Lessons
Three transferable takeaways came out of the project.
Persona Beats Adjective
"Be friendly" failed; "a six-year support specialist who writes like a calm human" succeeded. Describing a person carries far more signal than stacking adjectives.
Banned Phrases Are Underrated
Explicitly forbidding "reaching out" and "don't hesitate" did more for the voice than any positive instruction. Sometimes the fastest fix is telling the model what not to do.
Measure One Real Number
Tracking the no-edit send rate kept the team honest. Before measuring, they argued about whether the drafts "felt better." After, they had a number that settled it. Run any persona you build past our role prompting checklist for 2026 before deploying.
What the Team Tried Next
The support win prompted the agency to apply the same persona discipline elsewhere, with mixed results that are themselves instructive.
A Win in Proposals
They built a persona for first-draft client proposals — "a strategist who leads with the client's problem and ties every recommendation back to it." It cut proposal drafting time noticeably because the structure was consistent from the start. The lesson held: a behavioral persona on a subjective, voice-sensitive task pays off.
A Non-Win in Data Entry
They also tried a "meticulous data analyst" persona on a task that extracted figures from spreadsheets into summaries. It changed nothing measurable. The task was objective — the numbers were either pulled correctly or not — and no persona affected that. The team learned to screen tasks for subjectivity before reaching for a role, exactly the distinction our best practices that actually work emphasize.
How to Replicate the Result
The arc here transfers to any team whose AI output sounds wrong. The replication recipe is short.
The Steps That Generalized
- Pick one measurable proxy for quality before changing anything — the team used no-edit send rate.
- Describe a person, not adjectives — a tenured specialist beat "friendly."
- Forbid the specific tics that signal robotic output rather than only describing the ideal.
- Place persistent personas in the system message so they hold across every request.
- Re-measure and keep the change only if the number moves.
None of this required new technology — only the structured persona process and the discipline to measure. That combination is what turned a frustrating tool into one the team actually trusted.
Frequently Asked Questions
Why did "be friendly and warm" make the output worse?
Because adjectives give the model no concrete behavior to execute. "Friendly" got interpreted as generic cheerfulness with exclamation marks, which clashed with the brand's calm tone. Describing an actual persona — a tenured specialist who writes like a calm, capable human — carried the real signal the model needed.
How did the team fix the robotic openers and closers?
By explicitly banning the offending phrases — "reaching out" and "don't hesitate" — and requiring the reply to acknowledge the specific issue in the first sentence. Forbidding the boilerplate directly was more effective than any positive tone instruction. It forced the model to engage with the actual ticket instead of reaching for filler.
Did role prompting improve the accuracy of the replies?
No, and the team did not expect it to. Personas change tone and structure, not the model's knowledge. When a ticket required information the model did not have, it still could not answer correctly. The team handled that with a separate escalation path for agents, keeping persona work focused on voice.
Why place the persona in the system message?
Because it needed to hold across every ticket the assistant handled. A persona in the system message governs the whole interaction and resists drift, whereas one buried in a per-ticket message would lose influence. For a tool processing many requests under one consistent voice, the system message is the right home.
What single metric proved the change worked?
The share of AI drafts agents sent without edits, which rose from about 8 percent to roughly 55 percent. Tracking one concrete number ended the subjective debate about whether drafts "felt better" and gave the team clear evidence that the persona redesign delivered real time savings.
Key Takeaways
- A bare prompt produced accurate but lifeless support replies that agents rewrote nine times out of ten.
- "Be friendly" failed; a concrete persona — a tenured specialist who writes like a calm human — succeeded.
- Explicitly banning boilerplate phrases fixed the robotic tone faster than any positive instruction.
- The no-edit send rate rose from about 8 to 55 percent, while accuracy stayed flat as expected.
- Measuring one real number settled the "does it feel better" debate and proved the persona's value.