# Tone Matching Tone matching (also called register adaptation) is an [[Large Language Models (LLMs)|LLM]] capability dimension that measures how well a model adjusts its output style — formality, warmth, vocabulary, sentence rhythm — to a target register: casual, professional, executive, intimate, technical, playful, etc. It is distinct from raw reasoning depth and from instruction adherence; a model can be smart and obedient yet tonally tone-deaf. ## Why it matters - Final-mile quality for writing-heavy workflows: the difference between "technically correct" and "shippable" is almost always tone - The differentiator for ghostwriting, customer communication, voice notes, drafting in a personal style, and any context where the human signals authorship - Hard to evaluate with standardized benchmarks because it depends on the target register, which is subjective and audience-specific - Increasingly used as a comparison axis in third-party model reviews where reasoning benchmarks have saturated ## What it actually requires - **Vocabulary control** — picking words that fit the register without slipping into corporate-speak, AI-isms, or generic phrasing - **Sentence rhythm** — varying length and cadence to match how the target voice actually flows - **Implicature** — knowing what to say between the lines (what to imply rather than state) - **Register consistency** — sustaining the tone across the entire output, not just the opening - **Restraint** — knowing when to stop, when not to add hedges, when to skip the closing summary - **Voice ownership** — sounding like a specific human, not a generic version of the requested register ## Observed differences across frontier models (May 2026) From community evaluations: - [[Grok 4.3]] — strongest for informal, close-friend register; lands the casual voice better than peers - [[Claude Opus 4.7]] — strongest for formal, executive-level writing; reads "without filler" but can feel regimented - ChatGPT (GPT-5.x) — improved natural phrasing vs earlier versions; balanced default - [[Gemini 3]] Pro — competitive but rarely the standout on tone No single model wins on every register. The honest take: pick the model that best matches the *target* register for the task, not a blanket champion. ## How to improve tone matching in practice - Provide a [[My Voice Profile|voice profile]] — a structured description of vocabulary, banned phrases, sentence patterns, openings, closings - Show, don't tell: include 2–3 representative samples of the target voice in the prompt - Specify the audience and the relationship (close friend, executive, peer, customer) - Ask the model to identify its register before drafting, then write - Iterate by pointing to specific tonal misses, not vague "make it more X" - Use a humanizer pass (e.g., the `osk-writing-humanizer` skill) to strip AI-isms and tonal flatness from any draft ## Caveats - Tone matching evaluations are subjective and audience-bound — what reads as "natural" to one reader reads as "off" to another - Models drift across versions; a model that nailed your voice in one minor release may not in the next - Risk of over-fitting to a sampled voice (mimicry) at the expense of adaptability across formats ## Related - [[Large Language Models (LLMs)]] - [[Grok 4.3]] - [[Claude Opus 4.7]] - [[GPT-5.4]] - [[Gemini 3]] - [[AI Sycophancy]]