AI Bot Setup #

Your AI bot is the heart of the platform. It reads your customer’s messages, understands what they need, and responds naturally - like a skilled team member who never sleeps. This guide walks you through setting up your bot step by step.

The good news: You do not need any technical skills. If you can describe how you want a conversation to go, you can set up your bot.

Where this lives. Everything on this page is on your AI Agent’s AI Instructions step — the first one in the agent editor.

What Your AI Bot Can Do #

Once configured, your bot can:

• Answer questions about your products, services, pricing, and policies
• Book appointments and manage your calendar
• Qualify leads by asking the right questions and collecting contact details
• Connect to other tools like your CRM, order system, or booking platform
• Follow up automatically when conversations go quiet
• Hand off to your team when it encounters something it cannot handle

Each Agent has its own bot configuration, so you can set up different bots for different purposes - a support bot, a sales bot, an event bot, and so on - and reuse each one across every channel or Broadcast that needs it.

AI Providers (How the Bot Thinks) #

The platform uses the latest AI models to power your bot. The Pro tier runs on Anthropic Claude (Sonnet) — it excels at following instructions and understanding context. Most accounts also have Max, our faster, lower-cost in-house tier, which is the best value for almost everyone (see Why Max Is the Best Value), and Mini alongside it — the same in-house family on a lighter setup, for the lowest price per reply. You can pick the tier per Agent in the AI Quality setting. Until you pick a tier and save, the Agent runs on Pro and bills 1 credit per action, even if Max is available on your account. (The older Economy tier is being phased out — use Max instead.)

You do not need to choose an AI provider. You choose an AI Quality tier on the Agent (Pro, Max or Mini) and the platform runs the model behind that tier for you. You just focus on telling the bot what to do.

Quick Start: Auto-Setup via Onboarding #

If you are setting the app for the first time, the guided onboarding wizard can configure your bot automatically. During onboarding, you enter your website URL and the platform’s AI analyzes your site to:

• Generate bot instructions tailored to your business, including tone, goals, and conversation flow
• Create FAQs from your website content so the bot already knows your products, pricing, and policies
• Set up your company information based on what it finds on your site

This gives you a fully functional bot in minutes. You can always refine the auto-generated setup later using the steps below.

The Recommended Playbook Workflow #

You don’t have to (and shouldn’t) write the perfect playbook on day one. The most successful bots on the platform — including ones that have closed millions of dollars in revenue — follow the same lean iteration loop. Here’s the workflow that works:

1. Generate the first draft from a URL #

Create the Agent by pointing it at a website (your sales page, product page, landing page). The platform generates the initial playbook from the page content — this fills in the Agent’s AI Instructions automatically. Don’t over-edit yet — just glance through the generated fields, fix anything obviously wrong, and move on.

2. Refine with Optimize with AI #

Open the Agent and click Optimize with AI — on an Agent it sits in the header of the Try Out chat, and it’s there on every Agent. Instead of rewriting the playbook by hand, give Optimize the big-picture objective in plain language, for example:

“We’re running a Black Friday promo — 30% off our annual package until December 1st. Adjust the playbook to mention this and push for sign-ups before the deadline.”

The AI rewrites the entire playbook to match the new objective. When it finishes, you get a review screen before anything goes live: every proposed change is shown side by side — how the field reads now versus what the AI wants to change it to. Keep the changes you like and switch off the ones you don’t; suggested FAQ additions, edits and removals are listed separately, and each one is only applied if you explicitly accept it. Apply saves your selection, Reject all puts the playbook back exactly how it was, and closing the review keeps it waiting for you to decide later. On an Agent, what you apply is saved as a draft: test a few messages in the Try Out tab to make sure it feels right, then click Publish to live in the amber banner at the top of the editor to ship it. Most playbooks should go live around 20,000 characters — well under the limit.

3. Iterate from live conversations #

Once the bot is live, watch real chats come in. When a bot reply isn’t quite right:

1. Open that chat in the Chat Interface.
2. Use the chat → playground → fix → save loop documented in Chat Playground: thumbs-down the bad turn, edit the playbook in the side panel, restart, save.

Each fix adds or adjusts a small rule in the playbook. Over the next 1–2 weeks — usually after 20 to 50 live conversations — the playbook converges. After that, you’ll rarely touch it again. Some of the longest-running bots on the platform haven’t been edited in two years.

Keep Your Playbook Lean #

The character limits exist to be generous, not to be filled. Successful playbooks tend to land in the 10,000–30,000 character range — not the 80,000+ ceiling. A dense, focused playbook outperforms a sprawling one every time.

Why density matters (the employee analogy): Imagine training a new employee. If you give them one clear goal, they execute it well. Give them ten different goals, hundreds of rules for every possible message, and a 50-page binder of context — and they’ll deviate from the main goal, get confused, and make more mistakes. The AI behaves the same way.

Concrete guidance:

• One Agent = one clear goal. If you have very different goals (sales, support, recruitment, booking), build separate Agents for them. A specialist bot per goal outperforms a generalist bot.
• Don’t dump entire blog posts, sales decks, or white papers into the playbook. Pull out the facts the bot actually needs to close the deal. Everything else goes into FAQs (where the smart search only loads what’s relevant per message).
• Stop adding rules pre-emptively. Wait until a live conversation goes wrong, then add a rule for that specific case. Rules you imagine you’ll need rarely match the rules you actually need.
• Half the limit is plenty. If you’re hitting the character limit, that’s usually a signal the playbook has lost focus — not a signal to raise the limit.

Setting Up Your Bot (Step by Step) #

1. In the main sidebar, under AI Studio, click AI Agents.
2. Open the Agent you want to configure.
3. Click its AI Instructions tab.

Primary language sits at the top of the tab, above everything else. It’s the language your AI writes in by default, and the language it uses for the notes it writes you — human-handover alerts, tasks, suggested FAQs. It is not a lock: your bot still answers each contact in whatever language that contact writes in.

Pick a new one and it saves immediately — this field doesn’t wait for Publish to live the way the rest of the tab does. What it does not do is rewrite anything already written: your existing instructions and FAQs stay in the language they were written in, so if you want the agent to switch languages completely, retranslate those too.

Two script checks run on their own, nothing to switch on. An agent set to Chinese (Traditional) or Chinese (Hong Kong) has any stray Simplified character rewritten to its Traditional form before a reply goes out (请问 becomes 請問), while text that is already Traditional stays exactly as written. An agent writing Japanese has every reply checked for characters that only exist in Chinese (mainland simplifications such as 这, 现 and 们, and old forms Japanese no longer writes); if one slipped in, the reply is regenerated once with a correction and the clean version is sent. Ordinary kanji, including names such as 達也, are never touched. The Japanese check runs on the Max and Mini tiers, which is where this happened; it also covers follow-ups and comment replies.

If your Primary language is Chinese (Traditional) or Chinese (Hong Kong), every reply is checked before it goes out and any Simplified Chinese character the AI slipped in is rewritten to its Traditional form (请问 becomes 請問, 适合 becomes 適合). Text that is already Traditional is left exactly as written, including 台 in 台北 or 台灣 and the 里 in an address, so this only ever removes stray Simplified characters. It covers normal replies, follow-ups, comment replies and photo captions, and there is nothing to switch on.

The rest of the tab is organized into four collapsible sections — Persona & Goal, Business Context, Conversation Behavior, and Escalation & Wrap-up. If you used the onboarding wizard, many of these fields will already be filled in for you. (Step 5 below, adding FAQs and resources, happens on the separate FAQs & Knowledge tab.)

New Agents get a leaner field set. Every new Agent starts on the current format, which keeps each section to its core fields (below). Older Agents created before this change may still show a few extra fields — Source of the Contact (Persona & Goal), Unique Selling Points, Market Positioning, and Pricing Info (Business Context), and Objection Handling and Questions (Conversation Behavior). Those work exactly the same way if you see them — just use them for the extra context they ask for.

1. Persona & Goal #

Who the AI is and what it’s after.

Think of it like training a new employee: You would tell them their role, how to greet customers, what to recommend, and when to ask for help. Keep messages short — tell the bot to write like a real person texting, 2-3 sentences per message.

2. Business Context #

Facts about your company the AI can lean on.

3. Conversation Behavior #

How the AI should steer and guard the chat.

Example of a rule that also covers common objections:

Never give a final quote in chat — share only the published starting price.
If someone says "it's too expensive," mention the money-back guarantee and
compare it to the cost of not solving the problem, instead of dropping price.

4. Escalation & Wrap-up #

When to hand off to a human and when to stop.

When the bot alerts a human, the chat is paused and you receive a notification. The bot will not send any more messages until you resume it.

5. Add FAQs and Resources #

• FAQs give the bot precise answers to common questions. See the FAQ Management guide.
• Resources are links the bot can share during conversations (your website, pricing page, booking page, product catalog, etc.).
• Other information (in the Extra context section at the bottom of the FAQs & Knowledge tab) is a free-text field the AI reads on top of your instructions. Use it for anything temporary or swappable — a sale running this week, or industry-specific context for the niche you’re targeting — so your permanent instructions stay untouched.

Making Your Bot Feel Human (Response Settings) #

Further down the same AI Instructions tab, the Reply Timing and Response Limits sections control the pacing and style of your bot’s messages so they feel like they are coming from a real person, not a machine.

When to use Enforce max messages: By default the limit is a strong instruction to the AI, which it almost always respects — but occasionally a chatty reply slips through as extra bubbles. Turning this on guarantees the limit, at a small cost: combined messages can feel denser and less natural than the AI’s own pacing. Leave it off unless split replies are causing problems for you.

Don’t judge the pacing in Try Out. The Try Out test chat is deliberately sped up so you aren’t sitting there waiting: it skips the wait before the first message entirely and puts a flat few seconds between the split messages. It never uses the response delays above. Your live chats do. If you want to see the real pacing, look at a real conversation.

How it works: Instead of sending one long paragraph, the bot breaks its response into multiple short messages - just like a real person texting. It also pauses between messages and shows a “typing…” indicator (on WhatsApp Web) to make the conversation feel natural. You don’t switch any of this on: it’s automatic on every live reply. Setting Max messages per response to 1 or 2 is what turns it off in practice, since there’s nothing left to split. One exception: a very long answer (roughly 1,500 characters or more) is sent as a single message whatever the limit says.

Which value gets used? The bot does not pick a random number between the two. It looks at how long the chat has been quiet since its last message: a reply that comes within a minute gets the Min wait, and a chat that has been idle for 30 minutes or more gets the full Max wait, with everything in between scaled smoothly. So if you set a large Max (say 300 seconds), the first reply after a long gap will wait the full 5 minutes. That is expected, not a fault; lower Max if you want faster first replies after a lull.

If an Agent has no response delay configured at all, the bot falls back to a 30-60 second delay.

Why such a long default delay? Customers ask to lower the delay all the time. Don’t. Instant responses signal “bot” and prospects’ guard goes up. The 15-30 second window plus split messages plus a casual tone makes the AI feel human, which is what actually closes deals.

Stay close to the 15–30 second default. Don’t set the minimum to 0 or 1 — replies that come back instantly read as robotic, and the pause is what makes the AI feel like a human typing. If you really need snappier replies, drop the minimum to 5 seconds, but never zero it out. Going the other way hurts too: a minimum of a minute or more means a prospect who asked a quick question is left staring at nothing, and it widens the window in which their follow-up messages get merged into the same reply.

How Rapid-Message Handling Works #

If a contact fires off several short messages in a row (“hey”, “are you there?”, “I wanted to ask about pricing”), the bot waits for them to finish, batches all the messages together, and replies once with a single coherent response.

This is a feature, not a delay. It prevents the AI from interrupting itself mid-thought and stops it from sending three separate replies to what was really one question. The wait is part of why the conversation feels natural — a real person would also wait for the other side to finish typing before responding.

How long it waits is your response delay. Batching is not unlimited: the bot only merges messages that arrive before it starts replying, and that window is the Min/Max Response Delay you set above. With the default 15-30 seconds, a follow-up message sent 45 seconds later arrives after the reply has already gone out, so it gets its own separate answer. That is working as designed, not a fault. If contacts habitually think out loud across several messages and you want more of them merged into one reply, raise the response delay; if you want faster replies, expect fewer merges. You cannot have both.

AI Transparency (Telling Contacts They’re Chatting With AI) #

Many countries now require businesses to tell people when they’re chatting with an AI — in the EU this is the AI Act’s transparency rule, which applies from August 2, 2026. Your AI Connector handles this for you with the Tell contacts they’re chatting with AI switch, on your agent’s AI Abilities step. The setup wizard shows the same toggle while your first agent is being created, and whenever it’s on, the AI Instructions editors and the Try Out test chat carry a note with its own switch — so you can see it’s active, understand why the AI introduces itself as an AI, and turn it off right there without hunting for the setting.

When it’s on, the very first AI reply a contact ever receives naturally works in a brief mention that it’s an AI assistant — woven into the greeting, not a legal banner. That’s it: every later reply, including automatic follow-ups and any future conversation the same contact has with you, is completely unchanged, and the bot keeps its normal tone and pacing throughout. A one-line mention at the start doesn’t hurt a well-run conversation, and it keeps you on the right side of the rules.

A few things to know:

• New agents have it on by default. Agents that existed before this setting arrived are left exactly as they were — turn it on per agent whenever you’re ready.
• You can turn it off. If your audience isn’t covered by transparency rules, flip it off — the editor shows a short note reminding you that compliance for your audience is then your call.
• The switch goes live when you publish. Like every other change in the Agent editor, flipping this toggle creates a draft: the Try Out chat uses it straight away, but real conversations keep the old setting until you click Publish to live (the amber banner at the top of the editor reminds you). If the AI keeps introducing itself after you turned the toggle off, an unpublished draft is almost always why.
• With it on, the bot doesn’t pretend to be human when asked. If a contact asks “am I talking to a bot?”, the AI answers honestly rather than denying it, even if your playbook tells it to say otherwise. With the toggle off that guarantee goes away, so anything your playbook says about claiming to be human is on you. Either way, don’t instruct it to claim it’s human: that’s what transparency laws actually punish.
• If an instruction like “say you’re a staff member” seems ignored, this setting is why. While the toggle is on, instructions about presenting the AI as a human or staff member are deliberately overridden — that’s the setting doing its job, not the AI failing to follow your playbook. The AI Instructions editors show a note with its own switch whenever this is active, so you can flip it off on the spot (and the compliance note above applies).
• Opening messages you wrote yourself aren’t touched — they’re your words, not the AI’s, so no disclosure is added to them.

AI Speed Modes (How Deeply the Bot Thinks) #

AI speed modes control how much “thinking” the bot does before responding. Think of it like asking someone a question - sometimes a quick answer is fine, and sometimes you want them to really think it through.

Recommended: Leave it on Human Like unless you have a specific reason to change it. The accuracy gain is usually worth more than the slight extra latency.

AI Quality (Quality vs Cost) #

Separately from AI Speed, you can also pick the underlying AI model. On the AI Instructions tab this setting is labeled AI Quality. It’s purely a cost/quality knob.

Max runs entirely on our infrastructure — no third-party API keys to manage, no rate limits to worry about, and no surprise external bills — and it’s the most cost-efficient tier. It’s available on eligible accounts; for everyone else, the Standard (Pro) tier is the default.

Mini shows up right under Max on eligible accounts. It runs on the same infrastructure as Max at a lower per-action price — the trade is a small step down in accuracy on tricky factual questions. For the exact per-action prices of both, see the Billing System guide.

Pro stays available for users who specifically want the original Anthropic model with years of production miles behind it. Quality-wise Max and Pro are interchangeable for the vast majority of use cases. If you’re already on Pro and happy with it, there’s no reason to switch. You can change tiers anytime based on your cost/quality preference.

Economy is deprecated. It’s superseded by Max, which uses a newer-generation model. Existing Economy Agents are unaffected — they keep running on Economy until you change the tier — but the Economy option in the picker is greyed out unless the Agent is already on it. Once you switch an Agent to Pro or Max, you can’t switch back to Economy.

“Pro” here is an AI Quality tier, not a subscription plan. It only sets what one AI action costs; it has nothing to do with which plan you’re on.

For the exact credit cost of each tier, see the Billing System guide.

Bot Message Limits #

Every Agent has a Max AI Messages setting (in Response Limits) that caps how many AI replies the bot can send to a single contact before pausing:

• Default: 20 AI replies per contact
• Configurable to higher values, or to unlimited where your plan allows it
• When the cap is hit: The bot is paused for that contact. On a direct account you’ll get an email and push notification so a human can take over the conversation. Managed accounts: this alert follows your provider’s notification settings, so if they have those emails switched off, no email is sent.

This is a safety net so a single chatty contact cannot drain your credits indefinitely. Only AI replies count toward this limit, not the contact’s incoming messages.

Separately, some plans carry a hard limit on how many AI replies are even possible per chat. When that ceiling is reached, the bot simply stops replying to that contact — with no email. You can still message the contact manually.

Once the hard limit is reached for a specific contact, the bot stops responding automatically.

Bot Instruction Budget (How Much You Can Write) #

The Context Budget shown on the AI Instructions tab is a character budget you spread across most of your bot-instruction fields — persona, goal, company info, and so on. A 20K tier means roughly 20,000 characters total across those fields (about 3,000-4,000 words). This is about how much you can write in your instructions, not how much of the conversation the bot remembers — the bot always sees the recent conversation history regardless of tier, and very long chats are automatically summarised in the background at no extra cost.

Every instruction field in the four sections above counts toward the budget, including Rules and Conversation Flow. What doesn’t count is your knowledge base — FAQs, imported documents, product and pricing details all live separately and never touch this budget. That’s the right home for bulky, client-specific reference material: the AI pulls in only the entries relevant to each conversation, so your knowledge base can grow far beyond your character budget.

Which one you get depends on your plan. All three plans sold today — Business, Agency and Agency Unlimited — include the full 100,000 characters. On the AppSumo lifetime tiers the budget scales with the tier: 20,000 on Plan 1, 50,000 on Plan 2, 75,000 on Plan 3, and the full 100,000 from Plan 4 up. Your own figure is the one shown on the Context Budget meter.

For most Agents, 20,000 characters is plenty. As a reminder, the budgets are meant to be generous — a lean, focused playbook well under the limit outperforms one that fills it.

You don’t have to count for yourself. While you edit your instructions, a bar shows how much of your budget you’ve used (“12,340 / 50,000 characters used”). It turns amber as you get close to the limit and red once you’re nearly out, and it counts every instruction field together rather than each one separately.

How Much of the Conversation the AI Remembers #

The AI reads the recent conversation in full every time it replies — enough to cover almost every real conversation end to end.

Older messages aren’t thrown away either. Once a chat grows very long, the platform summarises the older part in the background and the AI keeps using those summaries alongside the recent messages, so it still knows what was agreed weeks ago. Mini runs with a lighter conversation memory than Max — that’s part of its lower price — and switching an Agent back to Max restores the full memory immediately.

Conversation length never changes what you pay: an AI reply costs the same flat credit rate whether the chat is 10 messages long or 1,000. There’s nothing to configure — memory isn’t a setting.

Editing the Playbook as One Block of Text #

Once you’ve written a playbook you like, you don’t have to retype it for the next Agent, and you don’t have to edit field-by-field if you’d rather work in a text editor. Click Edit as Text at the top of the AI Instructions tab to switch from the structured fields to one plain-text block, with each section under its own # heading:

#personality
Friendly, direct, never pushy...

#goal
Book a demo call...

That means you can also copy the whole block out, edit it in any text editor or document, then paste it back — handy for copying a playbook you like into a new Agent. Click Edit as Fields to switch back. Before anything saves, you’re shown exactly which sections would change — nothing is written until you confirm. Any heading that isn’t recognized is flagged rather than silently ignored, so a typo in a heading can’t quietly wipe a section.

How AI Credits Work #

The platform uses a simple credit system to track AI usage:

What uses credits - the rate is your Agent’s AI Quality tier (Pro 1 credit, Max 0.25, Mini 0.15, Economy 0.5, and 0 with your own Anthropic key except Max and Mini, which always bill 0.25 and 0.15):

• Each time the AI generates a response (one “thinking” cycle, even if it sends multiple message bubbles)
• Each time the AI uses a tool (like booking an appointment, searching your website, or calling a custom function)
• Each time the AI performs a background task (like extracting a contact’s name or evaluating the conversation)

What does NOT cost credits:

• Sending a message manually yourself
• Receiving messages from contacts
• Template messages (these have their own separate cost)

Credits included with your plan:

Your plan comes with a pot of AI credits. The Monthly Grant tile on Settings → Billing is the only authoritative figure for your own account.

You can purchase additional credits at any time or enable auto-recharge so your bot never runs out mid-conversation.

Bring Your Own Key (BYOK): BYOK lets you connect your own Anthropic API key (a private code from Anthropic that lets your account send requests straight to their AI service) so that AI replies are billed directly to your Anthropic account instead of consuming platform credits. If BYOK is enabled on your account, you can configure it from Settings → Advanced → BYOK API Keys. Channel infrastructure costs (like the WhatsApp Web monthly proxy fee or per-message WhatsApp delivery fees) still apply.

Testing Your Bot in Try Out #

Before going live, always test your bot in the Try Out tab. It simulates a real conversation so you can see exactly how your bot will behave.

How to use it:

1. Open the Agent.
2. Click its Try Out step — the last one in the editor.
3. Type a message in the input box at the bottom, as if you were a real customer.
4. Press Enter. The bot replies using the Agent’s full configuration (instructions, FAQs, tone, model, everything).
5. If a reply isn’t right: go back to the AI Instructions tab or the FAQs & Knowledge tab, make your changes, and click Save.
6. Return to Try Out and test again. Saved changes apply to the test chat right away — but they are not live yet: real conversations keep using your previous setup while you test.
7. Repeat until you are happy with the responses.
8. Click Publish to live in the amber banner at the top of the editor. That’s the moment your changes take effect for real customers. Changed your mind instead? Discard in the same banner resets everything back to the live setup.

What to test:

• Common questions your customers ask
• Edge cases (off-topic questions, gibberish, competitor mentions)
• Whether the bot correctly alerts a human when it should
• Appointment booking flows (if applicable)
• That the bot stays within the boundaries you set

Try Out uses the exact same AI as live conversations, so what you see in testing is what your contacts will experience.

Tips for Getting the Best Results #

1. Fill in everything. The more information you give the bot (company info, FAQs, resources), the better it performs. Do not leave sections empty.

2. Structure your instructions clearly. Use numbered lists and clear sections. The AI follows structured instructions much better than long paragraphs.

3. Add FAQs for common questions. If contacts keep asking the same things, add them as FAQs. The bot automatically finds the right FAQ for each question. See FAQ Management.

4. Use “Other information” for temporary or niche-specific knowledge. The field lives in the Extra context section of the FAQs & Knowledge tab. Running a sale this week, or targeting a specific industry? Add that context there so the bot knows without you changing your permanent instructions.

5. Add resource links. Include links to your website, pricing page, booking page, or product catalog. The bot shares these naturally during conversations.

6. Review conversations regularly. Check your live conversations to see how the bot is doing. If it handles something poorly, update your instructions or FAQs.

7. Use “Optimize with AI”. Click it in the header of the Try Out panel, describe what should improve (or thumbs-down a reply that missed the mark), and review the proposed rewrite side by side before accepting anything. It works from what you tell it, not from your past live conversations. For those, read your Daily Summaries and bring the takeaways here. See Optimize with AI.

8. Set up auto-tagging for key actions. Auto-tagging lives on the Agent’s own Tags step. Click + Add Tag and fill in two fields per tag: a Tag Name (e.g. booked) and a Tag Description and When to Assign (e.g. “Assign this tag when the contact confirms an appointment with a specific date and time”). The bot uses your description as the rule, and applies the tag automatically when a conversation matches. Tags you create here can then trigger webhooks (automatic notifications sent to another app or system) or deal stage changes elsewhere in the platform.

9. If you’re using AI-powered booking, instruct the bot to ask for the contact’s email BEFORE booking. Otherwise the Google Calendar invite has no one to send the confirmation to, and the contact never gets the meeting link. Add a step to your Conversation Flow like: “Before booking any appointment, ask the contact for their email address and confirm it back to them.”

10. Renaming something across the playbook? If you rename a product line, a pricing tier, or a team member, you need to update every reference to it across the playbook, instructions, FAQs, and resources. Two ways to handle this:

    • Switch to Edit as Text. Copy the playbook text out into a tool like Notes, Word, or VS Code and use find-and-replace. Paste it back when you’re done.
    • Use “Optimize with AI”. The AI will rewrite the playbook for you, but this costs credits — so save it for bigger restructures, not single-word renames.

11. Don’t dump entire blog content into the playbook or FAQs. It’s tempting to paste in every blog post, white paper, or sales deck you have, but bloated context hurts accuracy. The AI does best when the playbook is focused and the FAQs are crisp. If you have several very different goals — sales, support, booking, recruitment — build multiple Agents, rather than cramming one bot full of everything. Each specialist Agent will outperform a giant generalist.

12. If you write example messages, write them in every language you serve. Example phrases are the single strongest influence on how your bot writes — it copies their wording far more closely than it follows a written rule. So if all your examples are in one language, the bot drifts back into that language mid-conversation. This happens even when you’ve told it to match the contact’s language, and even after it has replied correctly for several messages. The giveaway is that it switches exactly at a step you wrote an example for (confirming details, offering meeting times), while the free-form parts stay in the right language. Either give each example in every language you support, or cut them down to short skeletons and add a line like “these examples are written in Catalan — always write in the contact’s language.” Do the same for any handover or fallback message you have written out.

13. Watch for instructions that fight a setting. Some things are controlled by a setting, not by your instructions, and the setting always wins. The most common one: writing “never split a response into two messages” while Max Chunks is set to 6. If you want single messages, set Max Chunks to 1 — see Response Settings.

Frequently Asked #

I edited the Agent — why hasn’t anything changed? Saving an edit doesn’t make it live anymore — it saves a draft that only the Try Out chat uses. When an unpublished draft exists, an amber banner appears at the top of the editor; click Publish to live there and the change takes effect everywhere the Agent is used right away.

Can two Agents share the same FAQ library but answer from different parts of it? Yes — FAQs live in your shared library and are linked per Agent, so each Agent can link its own set. See FAQ Management.

My Japanese agent put a Chinese character into a reply. Do I have to fix that myself? No. Japanese replies are checked before they go out and a reply carrying a Chinese-only character is regenerated once with a correction, on the Max and Mini tiers. Keep the agent’s Primary language on Japanese and write your example messages in Japanese too — the AI copies the language of its examples more closely than any rule you write. See Primary language above.

My Agent isn’t answering anyone yet — can I still test and improve it? Yes, all of it. The Try Out chat runs on the Agent itself, so a brand-new Agent can be tested the moment you’ve written its first instruction, and Optimize with AI and the per-reply thumbs-down work there too. What decides which real conversations reach the Agent is its Entry Points — see Entry Points.

Next Steps #

• AI Agents — what an Agent is and how the tabs fit together.
• Managing FAQs — give your Agent ready-made answers.
• Why Max Is the Best Value — the full case for the Max tier.
• AI Model & BYOK — bring your own Anthropic key.