Stop Your AI Agent from Writing Like a PR Bot

Most AI agents have a writing problem. Left to defaults, they reach for "game-changing," pad every sentence with a subordinate clause, and sign off like a LinkedIn influencer. That is fine for demos. It kills trust in production.

A small open-source project called Claudette (specifically the nobuzz prompt layer) surfaced recently as a minimal fix for Claude's tendency to default to BuzzFeed-style filler. The idea is simple: constrain the model's register at the system prompt level. The execution is instructive for anyone building agents that produce customer-facing text.

Why LLMs Default to Corporate Mush

Large language models are trained on the internet. The internet is full of press releases, content-farm articles, and LinkedIn posts written to sound impressive rather than communicate. The model learns that certain phrases pattern-match to "professional writing" and reaches for them by default.

The result: your agent writes "leverage" instead of "use," buries the point in a subordinate clause, and opens every email with "I hope this finds you well." None of this is hallucination. It is just the model doing what got rewarded in pretraining.

The Fix Is Not Magic, It Is Constraints

The nobuzz approach is essentially a denylist: a system prompt that names forbidden phrases and patterns, then reinforces a specific register (terse, direct, first-person). A few mechanics that actually work in production:

Explicit phrase bans. List the exact strings you do not want. "Do not use the word 'leverage'" works better than "write concisely." Vague instructions produce vague compliance.

Positive anchors, not just negatives. Pair the ban list with concrete examples of the voice you want. Two or three short before/after pairs in the system prompt calibrate the model faster than paragraphs of adjectives describing tone.

Structural rules beat adjectives. "One idea per sentence. No sentence longer than 20 words. Lead with the verb" produces more consistent output than "write in a clear and engaging style."

Evals catch regression. If the agent runs at scale, add a cheap LLM-as-judge eval that flags forbidden phrases in outputs. Without it, a prompt tweak elsewhere will quietly reintroduce the mush.

Where This Matters in Agent Pipelines

Voice drift is annoying in a chatbot. In a multi-step agent pipeline it compounds. Each node that generates text can introduce register inconsistency, and by the time the output reaches a user it sounds like it was written by a committee of robots.

The practical fix: centralize the voice constraints in a shared system prompt or a prompt template that every text-generation node inherits. Do not re-specify voice at each step. Do specify it once, strictly, at the top of the chain.

This also matters for cold outreach agents. If you are building an agent that drafts or personalizes emails, the default LLM voice will get those emails ignored. If you want a sense of how bad your current outreach copy is before you automate it at scale, our free Roast My Cold Outreach tool tears down a cold email and rewrites it in plain language. Worth checking before you wire an agent to send 500 versions of a bad template.

Practical Implementation Notes

A few things learned from building production agents that write customer-facing text:

• Temperature matters. Lower temperature (0.3-0.6) reduces creative drift but can make output feel mechanical. Find the floor where it stays on-register without sounding like a database query.
• The system prompt is not enough alone. Fine-tuning on domain-specific examples will always beat prompt engineering for high-volume, voice-sensitive workloads. Prompting is the right starting point; it is not the ceiling.
• Human review loops belong in the pipeline longer than you think. Ship with a human in the loop, measure where the agent drifts, then tighten constraints. Do not automate past the point you can observe.
• Version your system prompts like code. A voice regression from a careless prompt edit is hard to debug without a diff.

The Bigger Point

The nobuzz project is tiny. The problem it points at is not. Every production agent that generates text has an implicit voice policy, whether you wrote one or not. If you did not write it, the policy is "whatever the model defaults to," which is corporate filler optimized for sounding professional rather than being useful.

Taking 30 minutes to write an explicit, specific, example-backed voice constraint into your system prompt is one of the highest-leverage things you can do for user trust in an AI agent product. It is also one of the most skipped.

Build It Right From the Start

If you are building an agent that writes anything a customer or prospect will read, voice consistency is not a nice-to-have. If that is the kind of agent you want built properly, book a call and we can talk through the architecture.