You asked for output that is “professional but friendly, authoritative but humble, technical but accessible.” Each pair sounds reasonable. What came back is a five-paragraph essay that lurches sentence to sentence: one line reads like a McKinsey memo, the next like a friendly Slack message, the third like a textbook footnote. Re-prompting “blend them more smoothly” gives you a smoother but equally inconsistent voice.
The problem is not the blending. You asked for two voices, and the model averaged them. An average of two opposing voices reads as unstable, not balanced.
Fastest fix: name ONE primary tone, paste a 2-3 sentence example of exactly that voice, and demote the secondary tone to a short list of mechanical rules (for example “contractions allowed, no exclamation marks”). Adjectives like “warm” or “authoritative” are not instructions a model can reliably follow; an example is. The rest of this page explains why the “X but Y” pattern fails and gives you a step-by-step recipe.
Why this happens
When a prompt is not specific about voice, the model falls back to the most statistically safe style it learned in training — a “corporate helpful” default that many writers describe as Wikipedia-meets-LinkedIn. Two competing adjectives do not override that default; they just tell the model to oscillate between two thin caricatures of each word. Each adjective resolves to a different region of the training distribution, and “blend” gets interpreted as “switch.”
Which bucket are you in
| Symptom in your prompt | Root cause | The fix that works |
|---|---|---|
| Two adjectives joined by “but” / “and” (“formal but friendly”) | Model treats them as two targets and picks one per sentence | Pick one primary tone; demote the other to mechanical rules (Step 1) |
| Tone described in 1-2 sentences, no sample | Model has no anchor for what your blend sounds like | Paste a 2-3 sentence voice example (Step 2) |
| 4+ adjectives in the tone slot | Stakeholder politics, not a real voice | Cut to one primary; the rest become rules or get dropped (Step 1) |
| Domain genuinely mixes registers (sales, fundraising, conflict resolution) | Each section legitimately needs a different register | Write a per-section recipe (Step 3 and Step 6) |
| One short tone phrase, no structure | No recipe for the model to execute | Convert tone to measurable rules (Step 4) |
| Voice drifts only after several turns | In-message tone instructions get summarized away | Move tone to a system / project instruction (Step 5) |
Before you change anything
- List every tone adjective in your prompt.
- For each pair, ask “do these conflict?” If yes, you cannot satisfy both equally — one has to lead.
- Find one real sample of writing that hits the blend you want.
- Decide your one primary tone before re-prompting.
- Map any secondary tones to specific structural positions if relevant.
Information to collect
- The full tone instruction, as written.
- The output that mixed tones unevenly.
- One sample of writing that hits the blend you actually want.
- Model, temperature, and where the tone instruction lives (message, system prompt, or project instruction).
- Whether the inconsistency is sentence-by-sentence or paragraph-by-paragraph.
Shortest path to fix
Step 1: Pick one primary tone, demote the rest
Bad: "Be professional but friendly."
Good: "Primary tone: professional.
Friendliness shows up ONLY through: contractions ('we're', not 'we are'),
'you' instead of 'one', and no exclamation marks.
Everything else defaults to professional register."One winner; the secondary tone is reduced to a few mechanical rules the model can apply consistently.
Step 2: Anchor with a one-paragraph example
Adjectives are interpretation; an example is a target. Paste 2-3 sentences in exactly the voice you want:
Voice example (write in exactly this voice):
"Quick note on the rollout: we're holding launch until Friday. The
auth flow is failing on Safari, and we want it solid before customers
see it. Tracking the fix in INC-4123."A single example can read as an outlier; two or three establish a reliable template, so add a second sample if one doesn’t pin the voice (see “If it still fails”).
Step 3: Map nuances to structural positions
If you genuinely want two registers, give each one a place rather than asking for a per-sentence blend:
Section 1 (opening): warm, "we", contractions.
Section 2 (technical detail): formal, third person, no contractions.
Section 3 (call to action): warm, direct, second person.The model is far better at “use register X in section 2” than at “blend X and Y throughout.”
Step 4: Convert tones to measurable rules
Behavioral rules beat adjectives because they are checkable. “Two-sentence paragraphs” beats “punchy”; “lead with the counter-intuitive point” beats “engaging.”
Tone rules:
- Contractions: allowed
- Exclamation marks: forbidden
- Second person ("you"): required
- Industry jargon: max 1 per paragraph, defined on first use
- Average sentence length: 12-18 words
- Reading level: roughly 8th gradeMeasurable rules eliminate the per-sentence drift because each sentence either passes or fails.
Step 5: Pin tone where it persists
Tone written into a single chat message tends to get summarized away as the conversation grows, so the voice drifts after several turns. Move it to a layer that reloads every turn:
- ChatGPT: Custom Instructions, or a Project’s instructions (Projects sidebar -> your project -> “Instructions”).
- Claude: a Project’s custom instructions, or Claude Code’s
CLAUDE.mdfile, which is reloaded from disk on every run rather than carried in chat. - API: the
systemprompt, which is processed before every user message.
System-level instructions persist across all turns; in-message instructions do not.
Step 6: For inherently mixed-register tasks, write a recipe
Some tasks legitimately mix registers. A sales email needs warmth and authority. Don’t fight it with adjectives — give each beat a job:
Recipe:
- Open: warm, name the recipient, reference a recent shared context.
- Middle: shift to authoritative — concrete numbers, specific named cases.
- Close: warm again — invite the next step, no hard sell.A recipe is a structural anchor that adjectives cannot provide.
How to confirm it’s fixed
- Read it sentence to sentence: the output reads coherent, not stitched together.
- Blind-read test: hand the output to someone who has not seen the prompt. If they can describe a single coherent voice (rather than “it keeps switching”), it landed.
- Run the same prompt 3 times. Three outputs in a consistent voice means the tone is pinned, not random.
- Check the measurable rules from Step 4 against every sentence.
If it still fails
- The two tones may be genuinely incompatible. Cut one.
- Shrink the sample: have the model write 2 sentences first, lock the voice, then expand.
- Provide 2-3 examples instead of 1. Tone pins more reliably with more anchors.
- For brand work, treat tone as a config file: examples plus rules plus position mapping, all locked and reused.
- If the voice still defaults to “corporate helpful,” ban the tells explicitly: no em-dashes as connectors, no “in today’s fast-paced world,” no hedge words like “might” or “maybe.”
Prevention
- Default to one primary tone. Everything else is a mechanical rule or a positional nuance.
- Save tone anchors (one short paragraph per voice) for reuse.
- Never put “but” or “and” between two adjectives in a tone instruction.
- Audit production prompts quarterly for accumulated tone wishes.
- For team workflows, agree on one voice and one anchor, not “the brand vibe.”
FAQ
Why does “professional but friendly” produce worse output than just “professional”? Because the model reads the two words as two separate targets and satisfies one per sentence. “Professional” alone gives it a single register to hold. Adding a competing adjective without an example or rule just introduces oscillation.
Is it ever fine to use two tone adjectives? Yes, if you immediately ground them. Name one as primary and define the other as mechanical rules or section-by-section positions. The failure mode is two bare adjectives with no example and no structure.
How many voice examples should I give? Two or three. One can read as an outlier the model treats as a one-off; two or three establish the pattern. More than five rarely helps and eats context.
My tone instruction works at first but drifts after a few turns. Why?
In-message instructions get summarized as the conversation grows, so the tone rule fades. Move it to a persistent layer: ChatGPT Custom Instructions or Project instructions, a Claude Project or CLAUDE.md, or the API system prompt, which is reapplied every turn.
The output still sounds like generic AI even after I fixed the conflict. What now? That is the model’s default “corporate helpful” voice showing through. Replace adjectives with concrete behavioral rules (sentence length, reading level, banned phrases) and anchor with a real example. Explicitly forbid the common tells (em-dash overuse, “leverage” as a verb, hedge words).
Should tone live in the system prompt or temperature? Tone belongs in the prompt as rules and examples. Temperature controls randomness, not register — lowering it makes output more predictable but will not turn two conflicting adjectives into one coherent voice.