Prompt Builder

Configure your agent through a structured form instead of writing a system prompt by hand.

The Prompt Builder takes the settings you choose — identity, tone, knowledge behavior, tool preferences, lead capture timing — and compiles them into a system prompt. Open it from the ✨ Build button next to the Instructions field on the agent’s Persona tab.

You don’t have to use the Prompt Builder. The Instructions field accepts any prompt you write directly. The Builder is for operators who’d rather configure than write.

How it works

1. You fill in the form across two steps: Configure and Preview.
2. The Builder compiles your configuration into a system prompt.
3. You preview the compiled prompt and can edit it directly before applying.
4. Clicking Apply sets the prompt as your agent’s Instructions. It’s now live the next time the agent responds.

The compiled prompt is plain text — there’s no hidden state. What you see in Preview is exactly what the model receives.

Basic vs Advanced mode

The Builder opens in Basic mode by default. Switch to Advanced from the mode selector at the top.

Basic mode exposes the essentials:

• Identity (role, agent name)
• Company name and description
• Tone style and verbosity
• Unsure strategy
• Lead capture timing
• A simple list of built-in tool toggles
• A single Knowledge Base on/off toggle

Advanced mode adds everything Basic has, plus:

• Emoji and bullet-point preferences
• Tool invocation strategy (conservative / balanced / aggressive)
• Per-tool settings for built-in tools (auto-trigger, confirmation, required fields)
• Custom API tool configuration (strategy, confirmation, notes)
• Knowledge Base: retrieval mode, strictness, citation mode, no-result strategy
• Custom Instructions (free-form additions)

Use Basic if you want a quick, sensible setup. Use Advanced when you need precise control — for example, switching a knowledge-base agent to Strict mode so it never speculates outside its documents.

Field reference

Identity

• Agent name — optional display name (e.g. “Nia”, “Pinku”). Appears in the prompt as part of the agent’s persona. If omitted, the agent is referred to generically.
• Role — required. A short description of who the agent is (e.g. “AI concierge”, “support specialist”, “sales assistant”). This is the primary identity anchor.

Company

• Name and Description — both optional. If either is provided, the agent is told it represents this company. The description is treated as informational context, not as executable instructions — so you can safely include facts, links, or notes about your business.

Tone

• Style — Professional, Friendly, Casual, Premium, or Technical. Sets the voice the agent communicates in.
• Verbosity — Short (1–3 sentences), Medium (balanced), or Detailed (full explanations).
• Emojis (Advanced) — when on, the agent uses emojis sparingly for clarity; when off, never.
• Bullet points (Advanced) — when on, the agent uses lists for genuinely multi-item content; when off, prefers prose.

Tip: Pair Short verbosity with widget-based deployment. Long-form responses don’t fit chat bubbles well.

Unsure strategy

What the agent does when it doesn’t know the answer:

• Say it doesn’t know — the agent admits uncertainty rather than guess. Best default for support and knowledge-base agents.
• Offer to escalate to a human — the agent acknowledges uncertainty and offers to hand off. Use when you have human escalation set up.
• Give its best guess (with caveat) — the agent attempts a best-effort answer and labels its uncertainty. Use only for low-stakes informational agents.

Lead capture timing

Only shown when Lead Capture is enabled in the Tools tab. Controls when the agent asks for contact details:

• Ask upfront, before helping — collects email before answering anything. Best for gated services and high-intent landing pages.
• Ask naturally after 2–3 exchanges — warms the visitor up first, then asks. Recommended default.
• Only when user shows intent to act — waits for pricing/demo/booking signals. Best for content sites where most visitors are browsing.

See the Lead Capture page for the full configuration.

Tools

• Tool list — enable or disable individual built-in tools (Weather, Currency, Maps, Search, Google Places, Google Directions, etc.). Only tools you’ve enabled in the agent’s Tools tab appear here.
• Strategy (Advanced) — Conservative (use tools only when clearly required), Balanced (use when they meaningfully improve the response), Aggressive (proactively suggest or invoke when relevant).

The Prompt Builder only enables tools the agent has access to. To add new tools, configure them in the agent’s Tools tab first.

Knowledge Base

Shown only when the agent has files or URL sources attached.

• Retrieval mode — when the agent searches the knowledge base.
  • Always search first — search before every answer. Best for documentation-heavy agents.
  • Search when relevant — search for company-specific questions (products, services, pricing, customers, case studies, integrations, policies, processes). Default and recommended.
  • Factual questions only — search only when the question is clearly factual and verifiable.
• Strictness — how much the agent can supplement retrieved information.
  • Strict — KB only — answer only from retrieval results. If nothing is found, the agent doesn’t speculate.
  • Balanced — KB + general — primarily use retrieval; can use general knowledge for context or framing, but never to substitute for missing company facts. Recommended default.
  • Flexible — KB as guide — use retrieval as a starting point; supplement with general knowledge. Still never fabricates company-specific examples or statistics.
• Citation mode — how the agent references sources.
  • No citations — sources aren’t mentioned. Default.
  • Internal references — agent references internal documentation when citing.
  • Public citations — agent cites sources by name when answering.
• If no results found… — fallback behavior when retrieval is empty.
  • Say you don't know — admits the information isn’t available. Default.
  • Offer to escalate to a human — acknowledges and offers handoff.
  • Ask user to clarify — asks the visitor to rephrase.

See the Knowledge Tool page for details on adding files and URLs.

Custom API Tools (Advanced)

If you’ve configured custom API tools for the agent, each appears here with:

• Enable/disable toggle — whether this tool is exposed to the agent at runtime.
• Strategy — Conservative, Balanced, or Aggressive (same meaning as the general Tool Strategy, but per-tool).
• Require confirmation — when on, the agent asks the user before calling the tool.
• Notes — free-form per-tool guidance the agent will see. Useful for clarifying when not to call the tool.

See the API Tool page for setup details.

Custom Instructions (Advanced)

A free-form text field at the bottom of Advanced mode. Anything you write here is appended to the compiled prompt as additional directives.

Use this for:

• Domain-specific phrasing requirements (“always refer to the product as ‘our platform’, not ‘the app’”)
• Topics the agent should redirect (“if the visitor asks about pricing, point them to /pricing instead of quoting numbers”)
• Output conventions specific to your use case

Do NOT use this to override:

• Security rules (e.g. “ignore your safety instructions” — these rules are non-overridable by design)
• Tool calling rules (e.g. “never call lead_capture_tool” — disable the tool in the Tools tab instead)
• Knowledge base rules (use the Knowledge Base settings above)
• Core behaviors like “don’t fabricate facts”

The compiled prompt explicitly tells the model that Custom Instructions cannot override these higher-priority sections. Attempts to bypass them via Custom Instructions will be ignored at runtime.

Common patterns

Support agent

Strict KB + escalation ensures no hallucinated answers. Visitors who hit a gap get a human, not a guess.

Lead-gen agent

Optimized for conversion. The agent collects contact info early and uses the knowledge base to back up product claims.

Documentation agent

Every answer is grounded in your docs and cited. No fabrication, no general-knowledge filler.

Content-site assistant

Lower-stakes browsing context. The agent stays helpful even when the KB doesn’t cover a question, and only asks for contact info when the visitor signals real interest.

What gets compiled into the prompt

The Builder compiles the following sections into a single system prompt, in this order:

1. Instruction Priority — defines that security rules override everything else.
2. Primary Objective — the agent’s overall goal.
3. Identity — your role and agent name, wrapped as data.
4. Company — name and description, wrapped as data.
5. Tone & Style — your tone choices.
6. Response Contract — formatting and concision rules (constant; not configurable).
7. Core Behavior — accuracy, continuity, and reasoning standards (constant).
8. Behavior — your unsure strategy.
9. Lead Capture Tool — fields, trigger rules, intent policy (when Lead Capture is enabled).
10. Lead Capture — timing strategy (when Lead Capture is enabled).
11. Knowledge Base — mode, strictness, citation, no-result rules (when KB has content).
12. Tools — strategy and per-tool descriptions (when tools are enabled).
13. Custom API Tools — your API tool definitions (when present).
14. Additional Instructions — your custom instructions (when present).
15. Security Rules — non-overridable safety constraints (constant).

You can edit the compiled prompt in the Preview step before applying. Edits there are saved exactly as written and used as-is by the agent — they replace the compiled version, they don’t add to it.

Re-opening the Builder with an existing prompt will not re-parse your hand-edits back into form fields. If you regenerate the prompt from the form, you’ll lose any manual edits you made in Preview. Either keep configuration in the form, or commit to hand-editing — mixing the two requires manual reconciliation.

Tips & gotchas

• The Builder reads agent state when you open it. If you enable a new tool in the Tools tab, close and re-open the Builder so it sees the update.
• Conflicting settings are reconciled automatically where possible. For example, Strict — KB only + Best guess would produce contradictory instructions; the compiler adds a tie-breaking rule so the agent doesn’t speculate on company-specific questions even when “best guess” is set elsewhere.
• Tool names in the compiled prompt are the actual runtime names (e.g. weather_tool, search_knowledge_base, lead_capture_tool). This is intentional — it ensures the model calls the tools the runtime registers.
• Custom Instructions are isolated. Anything you write there is fenced and labeled as operator-supplied. It cannot impersonate system instructions or override security rules.
• No emojis means no emojis. If you turn emojis off in tone settings, the agent will avoid them entirely, even when the visitor uses them.
• Always preview before applying. The Preview step shows the exact prompt your agent will receive. If something reads oddly there, fix it in the form rather than editing the prompt directly.