Did you see the new Brand Voices update? Still feels half-baked.

Having spent the last quarter rigorously evaluating AI-assisted content platforms for our internal developer relations workflows, I’ve been closely monitoring Writesonic’s evolution, particularly its Brand Voices feature. The recent update, while adding measurable parameters like "creativity" and "formality" sliders, still falls critically short of delivering the deterministic, brand-safe output required for enterprise technical communication.

My primary issue is the lack of trainable, granular control. The system operates as a high-level stylistic filter rather than a true voice model. For instance, when I configure a voice intended for our Kubernetes tutorial documentation—requiring a specific balance of imperative mood, passive voice avoidance, and a defined terminology lexicon (e.g., "pod" vs. "container instance")—the output remains inconsistent. Benchmarking against a set of 50 pre-written sample paragraphs shows a variance in style adherence of over 40%. This is quantifiably worse than fine-tuning a base GPT model on a curated corpus, which we've achieved at a 12% variance in similar tests.

The new sliders introduce more variables but not more predictability. Increasing "formality" might indeed reduce contractions, but it doesn't enforce our required RFC 2119 key word usage ("MUST," "SHOULD," "MAY"). The "creativity" slider seems to just increase lexical variety, often at the cost of terminological accuracy—a non-starter for technical docs.

**Key observed shortcomings in the current implementation:**

* **No corpus-based training:** Inability to ingest a repository of past documentation to derive style rules.
* **No term blacklist/whitelist:** Critical for preventing trademark misuse or deprecated technical terms.
* **Inconsistent parameter persistence:** Voice settings appear to interact unpredictably with the "enhance" or "expand" commands, often overriding the configured tone.
* **Zero observability:** No logs or explanations on *why* a particular phrasing was chosen, making debugging impossible.

Until Writesonic provides an API-accessible method to inject a style guide, define syntactic rules, and offer detailed generation logs, this feature remains a cosmetic layer. For a community focused on DevOps and scalable systems, this lack of control and observability is a significant barrier. I'm curious if others have conducted similar stress tests and what your benchmarks for "voice consistency" look like. Have you found workarounds, or are we all waiting for a more engineering-centric approach from the platform?

—chris