Guides
A practical guide to writing advanced prompts for agent persona and tone, browsing Thoughtly's voice library, connecting BYOK, cloning custom voices, and optimizing speech for your caller population.
Last updated
Your AI voiceAI voiceAn artificially generated, natural-sounding voice produced by a TTS model. Thoughtly supports a library of AI voices and brand-specific cloning. agent's persona — the way it speaks, listens, and represents your brand — is the difference between a lead that stays on the line and one that hangs up in the first ten seconds. Thoughtly gives you two levers for this: the advanced prompt (which shapes personality and guardrails) and the voice library (which controls how the agent actually sounds). This guide walks through configuring both, from writing an effective persona prompt to browsing, cloning, and assigning voices that match your caller population.
If you're deploying agents for high-volume consumer lead conversion — insurance quotes, mortgage inquiries, education enrollment, healthcare scheduling — voice and persona choices directly impact contact rates and conversion. A clipped, robotic voice with no persona framing will tank your answer rate before the agent ever gets to qualify the lead.
The advanced prompt is the overarching instruction your agent considers before anything else. It lives in Settings → Advanced settings inside the Agent Builder. Think of it as the agent's character sheet — not a routing document.
The advanced prompt defines persona, tone, and high-level guardrails. It does not control call flow. Routing logic belongs in Speak nodes and Outcomes. If you put navigation instructions in the advanced prompt, the agent will inconsistently follow them and you'll spend hours debugging why it sometimes transfers and sometimes doesn't.
You are a friendly, knowledgeable mortgage intake specialist. Speak in short, confident sentences. Be warm but efficient — callers are often anxious about timelines. If a caller expresses confusion, summarize what you understand and ask one clarifying question. Never quote specific interest rates or fees. If asked about pricing, say "Your loan officer will walk through rates and options that fit your situation." Avoid giving tax, legal, or financial advice.Keep the prompt to 3–6 lines. Long prompts dilute the signal — the agent starts treating everything as equally important. After a few test calls, revisit the prompt to tune tone, not logic.
Advanced prompts can reference variables that are filled at runtime — contact attributes, automation payloads, CRMCRMThe system of record for leads, contacts, deals, and activity. Thoughtly reads from and writes to your CRM continuously. fields, or channel context. This lets you adapt persona behavior based on what you already know about the caller. For example, if the contact's state is known from a CRM field, the prompt can instruct the agent to acknowledge relevant state-specific information without hard-coding it.
Voice selection happens inside the Agent Builder. Click the Voice field in the right-hand sidebar to open the Voice Selector panel. The panel has two tabs: Saved (your workspace's saved voices, shown by default) and Explore (the full catalog from Cartesia and ElevenLabs).
The Saved tab shows voices your workspace has already saved, grouped by language. The active voice's language group is pinned to the top. Each voice row shows the avatar, name, platform (Cartesia or ElevenLabs), gender, and cost multiplier. Click Play to preview. Click the voice row to assign it to your agent.
The Explore tab gives you access to thousands of professional voices. The Expressive toggle is enabled by default, prioritizing low-latency voices optimized for real-time conversations. Toggle it off to browse the full catalog.
Filter chips let you narrow results:
Filters combine. For example, enable Expressive + Female + Spanish to find low-latency female voices in Spanish. You can also click the search icon to search by voice name or keyword.
Click anywhere on a voice row to save it to your workspace library and assign it to the current agent in one step. If you want to save a voice without assigning it yet, click the Bookmark icon. The voice appears in your Saved tab for future use across any agent.
If your Saved tab is empty after connecting BYOK, your ElevenLabs account may not have any cloned or generated voices yet. Use the Explore tab to save voices, or clone a new voice from the Saved tab.
Bring Your Own Key (BYOK) lets you connect your own ElevenLabs API key to Thoughtly. Instead of using shared platform credentials, your agents use your provider account directly. This gives you access to your personal voices, cloned voices, and provider-specific pricing.
A paid ElevenLabs plan (Starter or above) is required. Free ElevenLabs API keys cannot be connected because the validation step rejects them.
To disconnect, go to Settings → Integrations and remove the ElevenLabs connection. This removes all BYOK-linked voices from your workspace library. Agents using those voices fall back to the default voice. Cartesia voices are not affected.
When the public library doesn't have a voice that fits your brand, Thoughtly lets you clone a custom voice directly from the Saved tab. The Clone Voice modal supports both browser microphone recording and audio file upload.
If your workspace uses BYOK, cloned voices are created in your own provider account and tagged accordingly. This gives you full ownership and portability of your cloned voices.
Cloned voices are scoped to your workspace. They are private (not shared with other workspaces), team-wide (all workspace members can use them), and persistent until manually deleted. A single cloned voice can be assigned to unlimited agents simultaneously.
To remove a cloned voice: in the Saved tab, click the filled Bookmark icon to unsave it from your library. Deleting or unsaving a voice that is currently assigned to agents will cause those agents to fall back to the default voice.
Only clone voices from individuals who have explicitly authorized their voice to be used for AI synthesis. Keep written records of voice-use permissions.
After assigning a voice, tune it for your caller population. Thoughtly provides several controls under Settings → Presence and Voice Optimization.
Pick a voice accent that matches your caller's region. This improves trust and comprehension.
| Market | Recommended accent |
|---|---|
| North America | US English |
| Europe | UK English |
| ANZ | Australian English |
| Mexico | Mexican Spanish |
| Spain | Castilian Spanish |
| Argentina | Argentine Spanish |
| Brazil | Brazilian Portuguese |
| Portugal | European Portuguese |
Each agent has a Language dropdown in the conversation settings sidebar. Pick the language, then assign a voice in that same language. Mismatched voice/language pairs cause noticeable pronunciation problems. To serve callers in multiple languages, create a separate agent per language with a matching voice and Genius knowledge baseKnowledge baseA structured source of business information, FAQs, policies, product details, or procedures that an AI agent can use to answer accurately..
Cartesia voices have a Voice Speed slider in the Presence section of the Agent Builder sidebar. The slider runs from 0 (slowest) to 100 (fastest), with 50 as the default. Small adjustments of 10 points are noticeable. The slider is disabled when a non-Cartesia voice is active.
If your agent mispronounces names, brands, or industry terms, three options are available:
For numbers, IDs, or confirmation codes, enable Spell Numbers in the Speak node. This improves pronunciation of long numbers by reading them digit-by-digit.
Before deploying, use Thoughtly's built-in testing tools to validate that your persona prompt and voice selection sound right in real conversation.
Text-based tests are isolated from live calls. For voice quality, interruption behavior, pronunciation, and transfer timing, always place a real test call before going live. Test with native speakers in your target market when possible.
After deploying, track these metrics to evaluate whether your persona and voice configuration is working:
| Metric | What to look for | Where in Thoughtly |
|---|---|---|
| Answer rate (outbound) | Percentage of calls where the lead picks up. A jarring voice or off-putting persona drops this immediately. | Call history → filter by agent |
| Average call duration | Short calls (<30s) on inbound may indicate the persona is confusing or the voice is unpleasant. | Call history → duration column |
| Transfer rate | Percentage of calls that result in a warm transfer. A well-configured persona should transfer qualified leads, not alienate them. | Call history → outcome filter |
| Voicemail detection accuracy | If the agent is leaving messages on live answers, the voice or persona may be causing confusion. | Call history → disposition filter |
| Conversion rate by agent | Compare agents with different voices or persona prompts. A/B test by voice to find what converts. | Analytics → agent comparison |
Thoughtly's analytics dashboard provides call-level data. Review the first 50–100 calls after launch, listen to recordings, and adjust the advanced prompt, voice, or presence settings based on what you hear.
Yes. Each agent has its own Voice field in Settings. You can assign any saved voice to any agent, and a single voice can be assigned to unlimited agents simultaneously.
No. Cartesia voices are available regardless of BYOK status. However, connecting your own ElevenLabs key (BYOK) gives you access to your personal ElevenLabs voice library, cloned voices, and provider-specific pricing. A small number of enterprise workspaces use Thoughtly-managed credentials instead.
Yes. Voice cloning supports English, Spanish, French, German, Portuguese, Italian, Japanese, Korean, Chinese, Hindi, Dutch, Polish, Russian, Swedish, and Turkish.
Keep it to 3–6 lines. The advanced prompt defines persona, tone, and guardrails — not routing logic. Long prompts cause the agent to treat everything as equally important, which degrades performance.
Disconnecting removes all BYOK-linked voices from your workspace library. Agents that were using those voices fall back to the default voice. Cartesia voices are not affected. Reconnecting restores your ElevenLabs voices.
Thoughtly docs: Agent Voices — browse, preview, and assign voices
Thoughtly docs: Voice Optimization — accents, speed, and pronunciation
Thoughtly docs: Bring Your Own Key (BYOK) — connect your ElevenLabs account
Thoughtly docs: Voice Cloning — create custom voice clones
Thoughtly docs: Agent Settings — advanced prompt, presence, and post-call configuration
Thoughtly docs: Testing Voice Agents — Test Agent and Call Me tools
Thoughtly product: Voice Library — studio library, cloning, and A/B by voice
Thoughtly blog: How to Test and Iterate AI Voice Agents Before Going Live
Thoughtly blog: How to Use Thoughtly Variables for Dynamic Call Personalization
Thoughtly blog: How to Use Outcomes and Branching for Complex Call Flows