I've been evaluating DeepSeek Chat for several weeks now, primarily to assess its utility in my platform engineering workflow. One of my key use cases involves drafting customer-facing communications for incident postmortems, maintenance notifications, and cost anomaly explanations. While the generated content is factually accurate and structurally sound, I'm encountering a persistent issue: the output tone is consistently, and excessively, formal.
To illustrate, I prompted it to draft a notification for a scheduled Kubernetes control plane upgrade that would cause a brief API disruption. The result read like a legal document rather than a customer-focused update. For example:
> "Dear Valued Stakeholder, Please be advised that a mandatory infrastructure enhancement is scheduled for the Kubernetes control plane components. This procedural maintenance will necessitate a temporary cessation of API server availability. We anticipate a service interruption window not exceeding 300 seconds. We sincerely regret any inconvenience this may precipitate."
This is problematic for several reasons:
* It uses passive voice and Latinate vocabulary ("cessation," "precipitate," "advised").
* It distances the team from the action ("procedural maintenance" vs. "our team will be performing an upgrade").
* The formality can be perceived as cold or evasive, which undermines trust during operational events.
I've attempted to guide the model with explicit instructions like "use a collaborative and transparent tone" or "write informally," but it consistently reverts to this stilted register. My current workaround involves taking its technically accurate draft and manually rewriting all the phrasing, which negates much of the efficiency gain.
I'm curious if others in the community are facing this challenge, particularly those using it for SRE or customer success workflows. Have you developed effective prompting strategies or system prompts to coax a more human, direct tone from DeepSeek Chat for such communications?
My hypothesis is that its training data may be weighted toward academic papers and technical documentation, lacking sufficient examples of effective, modern operational communication. For comparison, when I feed the same scenario into a model fine-tuned on platform engineering docs (like Anthropic's Claude with a relevant context), the output is markedly more direct.
—Chris
Data over dogma
No, you're not the only one. It defaults to corporate robot because that's the safest, most common training data.
Trick is to write the prompt like you're briefing a junior engineer. "Draft a short, direct email for customers about a k8s API outage next Tuesday. Keep it plain, use active voice, no legalese. Acknowledge it's annoying but we're doing it for stability."
If you don't explicitly strip out the formality, it'll give you boilerplate. It has no inherent sense of what a "human" tone is, only what statistically follows your request.
Prove it.