Playground and AI setup

Where you try Yaplet before your visitors meet it — a walkthrough from an empty brand to ready to install, what the test chat does and does not do, and the AI setup that fills a brand in from your website.

What the Playground is for

The Playground is where you talk to your own Yaplet before anyone else does. You ask it questions the way a visitor would, you see what it answers from your knowledge, and you can change its instructions and its AI model and watch the effect straight away, before you save anything. Your visitors keep getting the saved version until you press Save changes.

That makes it the cheapest place to find a mistake. A wrong price, a rambling answer or a missing fact costs you nothing here. On your website it costs you a customer's trust.

Where to find it

Brand → (your brand) → Playground is the first tab of a brand, at /dashboard/brands/<brand id>/playground. Opening a brand without naming a page — a typed address or an old bookmark — lands on the first tab you are allowed to open, which for most people is this one. (A brand is the company or website Yaplet speaks for.)

A brand no longer has an Overview page. Its old address forwards here, and the things that page used to show now live in other places, listed under What happened to the brand Overview below.

You need the AI agent permission to open the Playground. Running AI setup from Brand settings or the chat widget needs the Brands & chat widgets permission.

The page has two halves: your settings in a column on the left, the test chat on the right. Where the brand has phone agents, a Chat / Call switch above the test area also lets you test a phone call (see Testing a phone call below). On a narrower screen a switch at the top, Test and Settings, shows one at a time.

From an empty brand to ready to install

Step 1 — Give Yaplet something to learn from

Yaplet answers only from what the brand knows, so the first job is making sure it knows something. Which of these looks like your screen?

  • You just finished onboarding, or you made the brand with + New brand. The setup is still running. Where the chat will be, you see its progress, and the instructions box is locked. Wait for it to finish; the test chat appears by itself. See Your first minutes: onboarding. If you chose I don't have a website yet — set up by chatting in onboarding, you ended up in a chat with Yaplet instead of here; open Brand → Playground from the menu whenever you want to test.
  • The brand exists but nothing has been added. The status card says Trained, but the small line under it reads No knowledge added yet. Yaplet has nothing to answer from, so all it can do is say it does not know. Press the Knowledge card, which opens Brand → Knowledge, and add your website, files, text or question-and-answer pairs. See Knowledge. Then come back.
  • The page says "This brand has no AI agent yet". The brand has no Yaplet answering for it, so there is nothing to test. Press Set up the AI agent. It opens the AI agent page, where one button, Add an AI agent, creates it already switched on. See Yaplet.

Step 2 — Read the status card

The card at the top of the settings column tells you whether Yaplet is ready to be tested.

What the card saysWhat it meansWhat to do
Building from your website…The setup you started in onboarding is still running.Wait. The chat appears when it is done.
Learning your website · 12/25 pages, with a progress barYaplet is reading the first 25 pages of your site. A note above the chat says "Still learning your website — answers get better in a moment."You can already ask questions, but give it a moment for better answers.
Writing test questions…The first pages are in and Yaplet is writing questions from them.Wait a moment.
Trained, and a small line such as Last trained 2 hours ago · 12 sources · yourshop.comYaplet has something to answer from. The line says when it last learned something, how many sources the brand holds and your website address.Start testing.
Trained, and No knowledge added yetThe brand holds no knowledge.Add some under Knowledge (step 1).
Could not read your sources, and Knowledge shows whyNone of the sources could be read and no website page has been learned.Open Brand → Knowledge to see why.
A green Tested badgeYou have asked the test chat at least one question.Nothing. It is the same as ticking Test your AI agent on the Getting started card.

Step 3 — Ask the test questions

Once Yaplet has learned most of your website's pages, a Test questions list appears under the status card: questions written from your own site by AI setup. Work through it like this:

  1. Press a question to open it. You see Answer on your website — what your site says — and a link to the page it comes from, which opens in a new tab.
  2. Press the send button at the right end of the row. It shows when you point at the row, and always on a touch screen. The question goes into the test chat as if a visitor had typed it (on a narrow screen the page switches to Test).
  3. Compare Yaplet's reply with Answer on your website. Same facts means your knowledge works. Different or invented facts mean you have something to fix (see When an answer is wrong below).

A brand that has never had its website read by AI setup has no list, so write your own questions. A good first round is:

  • Three questions you can answer in your sleep: opening hours, the price of your best seller, your return policy.
  • One Yaplet should not be able to answer, for example a product you do not sell. It should decline or hand over, never guess.
  • The same question in another language your visitors write in.
  • A vague one, like "it doesn't work".
  • "I want to talk to a person."

The test chat is drawn like your own widget, in its colors and corners. It opens with a fixed greeting ("Hi! What can I help you with?", not the greeting on your widget's Home screen) and your first three conversation starters as buttons, or three general questions if you have fewer than two starters. Type in the Message… box and press Enter. Like a real chat, it remembers what was said, so a follow-up can refer to the last answer. The New test chat button (the round arrow in the chat header) starts a clean one. When you come back to the Playground, the brand's latest test chat is still there; it is kept per brand, not per person. The round launcher under the chat opens and closes it, as on a website.

Step 4 — Change the instructions and see the effect before saving

The Instructions box is the text Yaplet follows in every answer: tone, house rules, facts it should always know. It is the same text as Answer instructions on AI agent → Answers.

  1. Write your own text, or press Start from a template and pick one of four starting points: Customer support (clear, friendly answers from your website, a person for account-specific help), Online store (finding products, shipping, delivery and returns, guiding hesitant shoppers to checkout), Sales and leads (a warm welcome, product and pricing answers, inviting leads to leave an email or book a call) and Bookings (services, availability, pricing and preparation, guiding visitors to book). A template replaces whatever is in the box, so edit it to sound like you. Nothing is saved yet, and the Undo changes button next to the menu puts back the last saved text.
  2. While the box differs from the saved text, the line under it turns into a warning: "Not saved yet — the test chat is already answering with these instructions." That is the point of the Playground: the next question you ask is answered with the new text.
  3. Ask the question that gave you a bad answer. For a fair comparison press New test chat first, so earlier answers in the conversation do not influence the new one.
  4. Happy with it? Press Save changes in the title bar. Not happy? Press Discard (or Undo changes). The title bar shows Unsaved changes while something is waiting, and leaving the page with unsaved changes asks first.

What makes good instructions is covered in AI agent settings. The short version: be specific, list what Yaplet should never do, and keep it short.

Instructions shape how Yaplet says things and what it avoids. They do not give it new facts: it still states facts about your business only from the brand's own knowledge. If the facts are wrong, fix the knowledge, not the instructions.

Step 5 — Choose the AI model

The Model row shows which AI model writes Yaplet's answers.

  • Where your plan lets you choose, a list opens. Pick a model and ask again: the test chat uses it straight away, before you save, so you can try the same hard question with two models. Save changes keeps your choice. Pricier models use more credits per answer, and the AI Models page shows how they compare.
  • Where it does not — the free 14 days and the free plan — the row shows the model we chose, locked, with "Upgrade for more advanced models" and an Upgrade button that opens the plans.

The model is one setting for the whole workspace, not just this brand. It is the same as the main model on Settings → AI Models and as Answer model on AI agent → Answers, so changing it here changes it everywhere that uses the main model. See AI Models. Your own role also has to allow changing models, not only your plan.

Step 6 — Check what it does when it cannot answer

Ask something your knowledge does not cover. What happens depends on When Yaplet has no answer, under AI agent → Human hand-off:

  • Politely decline — Yaplet says it has no information on that. It never guesses.
  • Connect to a person — on your website the visitor would be handed to your team. In the test chat you get a card instead, described under What the test chat does, and what it does not below.

The Actions row shows how many workflows and API tools are switched on ("5 actions enabled", with "2 workflows · 3 API tools" under it) and links to Brand → Workflows. With none it says Add your first action.

Step 7 — Install when you are ready

You are ready when:

  • The status card says Trained and lists your sources.
  • The test questions come back with the facts your site gives.
  • A question Yaplet cannot answer gets the reaction you want.
  • The tone sounds like you.
  • You have pressed Save changes. Visitors only ever get saved instructions and a saved model.
  • Under AI agent → General, Enable Yaplet is on and Website chat is on under Answers on. If Website chat is off, a note above the test chat says visitors on your website will not get these answers; the test works anyway.

Then press Install on website in the title bar. It opens the Install tab of your chat widget, with the code for your site (see Chat widget and Getting started with the widget). The button is the highlighted one when nothing is waiting to be saved, and an outline button while you have unsaved changes, as a reminder to save first. If the brand has no chat widget yet, it opens the Chat widget page, which offers to add one.

When an answer is wrong

Work out what kind of wrong it is, because each has its own fix.

What you seeUsually meansWhere to fix it
A fact is wrong, outdated or inventedThe knowledge is wrong, or missing the right pageBrand → Knowledge: correct or add the source. See How Yaplet finds answers.
Yaplet says it does not know, but your site has the answerThe page was not learned, failed to read, or is worded very differentlyBrand → Knowledge: check the source's status. See Why isn't Yaplet answering from a source I added?.
The facts are right but the tone, length or style is notThe instructionsInstructions, right here (step 4).
It hands over, or declines, when you wanted the otherThe no-answer settingAI agent → Human hand-off. The hand-off card links to it.
"The test answer failed. Try again." in red under the chatA temporary failureAsk again.

What the test chat does, and what it does not

The test chat is the real Yaplet: the same knowledge search, the same model, the same rules. What is switched off is everything that would reach your team or your customers.

In the test chatOn your website
Who sees the conversationOnly the Playground. It never shows up in your inbox, in the recent conversations on Overview or in your reports, and nobody gets a notification.Your team, in the inbox, with notifications.
Instructions and modelThe ones in the box and the list, even if unsaved.Only the saved ones.
Handing over to a personA card says "On your website, Yaplet would now connect this visitor to your team." Nobody is alerted. It links to Hand-off settings.The chat is passed to your team, who are notified.
API toolsThey run for real: Yaplet calls your actual system. A tool that changes data changes real data.The same.
Tools set to "Only for identified visitors"Refused. The test visitor is anonymous, so Yaplet passes on the message under How visitors get identified (yours, or the standard sentence if you wrote none).They work for visitors your website has identified.
Details a tool needs from the visitorYaplet asks for them in the chat text.The widget shows a form.
WorkflowsNot started. Yaplet picks a workflow only in a real website chat.They start as you built them.
The "Answers on" switchesA test works even if Website chat is off.If it is off, messages go straight to your team.
Per-visitor answer limits and free monthly answersNot counted.Counted.
The visitorA made-up visitor named Playground test, with your own email address.The real visitor.
Blocked wordsA message with one of your workspace's blocked words is refused.The same.
A tool that changes data (a refund, a cancellation, a booking) really runs from the test chat. Test it with a harmless record, or point it at a test system first.

Because workflows and identified-only tools cannot be tried here, test those on a real page. See Test your chatbots before going live.

Testing a phone call

If your brand has a phone agent, the Playground can test it too. A Chat / Call switch sits above the test area: Chat is the test chat described above, and Call lets you talk to the brand's phone agent from your browser. You don't need a phone number, and it works while the agent is paused. See Voice Agents for what a phone agent is.

  • Who sees the switch. People with the Voice & calls permission, and owners whose plan has no phone. For those owners Call carries a lock and shows a card, "Yaplet can answer your phone too", with a See plans button that opens the plans. Everyone else sees only the test chat.
  • No phone agent yet. With phone on your plan but no phone agent for this brand, Call shows "No phone agent yet" and a Create a phone agent button.
  • Which agent you call. The one a phone number rings, or the first one if none is rung. With several phone agents, a picker at the top of the left column chooses another one. It is locked while you have unsaved changes or a call is running.
  • Making the call. Press Start call, allow your microphone and talk. What you and Yaplet say, and what it looked up, shows as you speak. Use headphones, or Yaplet hears itself through your speakers. Chat can't be picked while a call is connecting or running, so hang up first.
  • The phone agent's settings. While Call is open, the left column shows the phone agent's main settings instead of the chat AI's: whether calls reach it (Answering calls, Paused or No phone number yet), Voice and language with a play button that lets you hear the greeting, the brand's Knowledge and Actions, the Greeting and the Instructions. The link at the bottom, All settings of this agent, opens the agent's own page for phone numbers, call recording, what to do when a caller asks for a person, pausing and past calls.
  • Saving. What you change there joins the title bar's Save changes and Discard, together with the chat side's instructions and model. A test call always uses the saved agent, so save before you call; the call panel reminds you while something is unsaved. Saving a new voice, language or greeting records the greeting again.
A test call is billed to your workspace like a real phone call, so it needs a paid plan, and it stops after 5 minutes. Actions run for real, as in the test chat.

Like the test chat, a test call stays out of your inbox. It shows in the agent's Calls list with a Test badge.

Playground or widget preview?

Two places let you try your chat, and they test different things.

PlaygroundLive preview on Brand → Chat widget
For testingWhat Yaplet says: knowledge, instructions, modelHow the widget looks and feels
What you seeA test chat drawn like your widget (colors, corners, launcher)The real widget in a frame beside the settings, with its Home and Chat screens
Unsaved changesInstructions and model count before savingLook, text and layout changes show before saving
The conversationStays out of the inbox and reportsA real conversation: it appears in your inbox, and its AI answer is paid for like any visitor's
Workflows, forms, hand-offNot started; a card instead of a hand-offWork as they do for visitors

The Test page button above the preview opens your last saved widget on a test page in a new tab. Use it for a last look before you tell anyone.

Who pays for test questions

A test answer is AI work like any other. It is paid first from the free 14 days' credits, then from credit you bought, then from your paid plan's usage. It never uses the free monthly answers that some older free workspaces still have. If nothing can pay, the question is not answered and the payment window opens. There is no daily limit, but if you send questions very fast, Yaplet asks you to wait a moment. See Your free 14 days and Plans and pricing.

When you arrive from onboarding

Onboarding ends here. Until the setup run has read your site, the middle of the page shows its progress instead of the chat, and the instructions box is locked. When it finishes, the test chat appears, and if Yaplet had no instructions yet, the description you typed in onboarding becomes its instructions.

Yaplet learns your website in two steps. The first 25 pages are learned quickly, so that test questions can be written from them; more follow in the background; after that Yaplet checks your site every hour, up to 2,000 pages. The status card shows the first step as Learning your website · 12/25 pages. The rest of the way is on Website pages.

What happened to the brand Overview

  • The setup checklist is the Getting started card in the sidebar, shown to owners. One of its steps, Test your AI agent, opens the Playground.
  • How Yaplet is answering — Questions, Answer rate and Handed to a person, for the last 30 days — sits at the top of every section of the AI agent page. The full picture is on AI Conversation Reports.
  • Trends for the whole workspace are on Overview in the main menu.
  • Sources and products show their own state on Brand → Knowledge and Brand → Products.

AI setup

AI setup reads your website and fills the brand in. You give it your website address — and optionally a sentence or two about the business — and it does the rest.

Where it runs

  • Onboarding — the first time the owner signs in, and for every brand you add with + New brand (that dialog asks only for a name, then continues with the same steps). It asks for your website and what Yaplet should do, then fills in the brand, designs the widget, learns the site and opens your Playground.
  • Brand settings → Fill from website — fills in the brand details. A tick box, Also read my website into the AI's knowledge, adds the website knowledge too.
  • Chat widget → Style → Redesign with AI — designs the widget only, and puts the result into the unsaved form so you can look at it before saving.

What it fills in

  • Brand details — company name, company description, website, the brand's language, the brand icon and the brand logo. The same company name, description and website are copied onto the Yaplet bot so it answers as that company.
  • Widget design — the Widget color and up to three Brand colors, Buttons & messages, the Home screen background (Color, Glow, Lines and Reaches), Cards, Corners, the launcher bubble's shape and Highlight ring (never which icon the launcher shows — that stays your choice), the Logo and Widget icon, and the widget's interface language. Replacing the design clears any colors set on the bubble itself, so the bubble follows your new brand colors.
  • Quick buttons — two to four questions a visitor of your business would ask, shown under the widget's Ask me anything... box. They only take the place of Yaplet's stock quick buttons: a button you wrote is never changed, and if you have already removed or rewritten all the stock ones, none are added. The box carries at most four buttons in total. They are written in the brand's language and translated into the widget's other languages in the background — until that finishes, visitors read them in the brand's language. Redesign with AI puts them into the unsaved form instead, and they are translated after you save.
  • Website knowledge — it learns your whole website, as described above, not a hand-picked set of pages. If nothing can pay for it, this part is skipped.
  • Opening questions — written from the pages it has just learned, and only when the Yaplet bot has none at all. If you wrote your own, or deleted ours on purpose, a re-run never puts them back — and a run that skips the website-knowledge part never adds them in the first place.

What it never touches

The chat widget's name, the brand's own name, and the public <name>.yaplet.help web address. Nothing you have already pointed an embed code or a saved link at is renamed underneath you.

It also never writes the widget's greeting or welcome line, and never picks a Home screen Photo. Whatever the Home screen tab holds stays as it is; where you have written nothing, visitors keep seeing Yaplet's built-in wording in their own language.

Your own screenshots

Wherever a run designs the chat widget, you can add Screenshots of your homepage (optional, up to 3): in the first step of onboarding, behind "Site behind a cookie banner or not live yet? Add a homepage screenshot", and in Redesign with AI. Yaplet visits your site and takes its own picture of it. It closes most cookie banners and pop-ups before taking the picture, but one can still get in the way, as can a slow-loading page — your screenshots take the place of that picture. The site is still read for its text, language, icon and logo, and if it cannot be reached at all, for example because it is not live yet, the run carries on from your screenshots.

  • PNG, JPEG or WebP, up to 8 MB each. A file that breaks a rule is left out on the spot, with a note naming it, and only the first three are used.
  • The field is hidden while Widget design is set to Keep what's there — the pictures would change nothing.
  • Fill from website does not offer it; it does not design the widget.
  • Screenshots are used for that one run and never stored.

When a run designed the widget from Yaplet's own picture, the result screen shows that picture with the caption Your widget is based on this view of your site. If something covers the page in it, the link Doesn't look right? Upload your own screenshots takes you back to the first step. That second run redoes only the widget design, not the brand details or the website knowledge. The picture is not stored.

Keep or Replace on a re-run

When the dialog opens, Yaplet checks what already has content and shows Keep what's there or Replace it underneath the address box, for exactly those parts — and stays quiet about the rest, which is why the very first run asks nothing at all.

  • Brand details default to Keep what's there: only empty fields get filled in, so a description you wrote by hand survives.
  • Widget design also defaults to Keep what's there, except in Redesign with AI, where a new design is the whole point and it defaults to Replace it: new colors, Home background, corners, card style and launcher look. Keep what's there leaves your colors, Home background, corners and launcher exactly as they are, and only fills in a missing logo or widget icon.

Quick buttons follow their own rule whichever you pick. A widget still in the look every new widget starts with counts as never designed, so a first design lands without anyone having to choose; once any of its colors, Home background, cards, corners or launcher have been changed, you get the question.

Who pays for AI setup

AI setup is paid like any other AI work: first from the free 14 days' credits, then from credit you bought, then from your paid plan's usage. There is no separate allowance of runs. If nothing can pay, the run does not start, you see "AI setup needs credit or a plan." and the payment window opens.

© 2026 Yaplet. All rights reserved.