The method: building an agent that holds up

The questions to ask

Depending on the agent's job, one more question, without which it won't hold up:

• support: where does the agent stop, and who does it hand off to? Without a boundary, it makes up an answer rather than admit it doesn't know;
• sales: what exactly do you sell, and what price can it commit to? It is the one thing a sales agent must never approximate;
• interview (the agent gets someone talking and writes up a report): what do you want to be able to measure at the end? With no number to collect, you get a nice story you can't use;
• internal: are team members' answers named or anonymous? Promising anonymity and then showing names destroys the whole team's trust.

Goals

Goals are not a formality: they decide what the console puts first. Each one shows in the “Chatbot goals” box on the agent page, with its trend and the number of conversations it is based on, and becomes a quick filter on the Conversations screen. Three at most, picked from a closed list:

• Each goal has a direction, up or down; a numeric target is optional, on purpose: an invented threshold does not survive the question “why 70%?” in a meeting.
• A goal specific to the business (“answer order questions on its own”) is set on a topic: first describe the conversation topics, with what belongs in each and what doesn't, then track the questions on that topic the agent resolves alone.
• With no history, start with two obvious goals. After a few weeks, have them suggested from real conversations: only those with actual data for this agent are kept.
• Whatever the goals, the Conversations screen keeps two filters for every agent: “Error” and “No answer”.

Entries that get found

• One question, one entry. An entry that groups ten questions under one heading resembles none of them: the right answer never comes up, even when it is written word for word. One FAQ line is one entry, one product is one entry.
• The visitor's words. They type “how do I pay”, not “payment methods”. A line in the entry with their real phrasings helps search find it.
• Statements, not instructions. The agent reads a retrieved excerpt as an instruction: write “delivery within 48 hours in the metro area”, not “never promise a lead time”. Instructions belong in the prompt.
• The source in the same sentence as the fact. The agent can then say where a number comes from instead of just asserting it.
• A short entry for each commercial answer (price, trial, lead time, scope). Buried in a long entry, a price lands in an excerpt that search never brings up. These entries are reread whenever the offer changes: the agent faithfully recites an outdated offer.
• Links that work. An agent that quotes a dead page does worse than quoting nothing.
• A site that contradicts itself makes an agent that contradicts itself (two prices for the same offer on two pages). Don't settle it silently: leave out the wrong page and ask the client to fix their site.

Checking it finds things

Information being in the base doesn't mean it gets found. Check with the client's real questions, rephrased the way a visitor types them, never with the entry's title: asking an entry for its own title proves nothing.

• Start with the question that prompted the project (the client's complaint, the most frequent case): the right entry must come up, with its link.
• In Test, turn on “Response details”: for each reply, what the agent read, the actions it ran, the role it used and how long it took.
• A wrong fact is tracked down everywhere before it is fixed: in the base, but also in the prompt, the welcome and the preset replies. Fixing only the base leaves the prompt contradicting it.
• An entry is used as soon as it is saved, without publishing a new version.

“Visitor memory”, in the same tab, remembers from one visit to the next what a person said. It is off by default: only turn it on if the client asks. It is tied to the visitor's identifier, so two people sharing one identifier share one memory.

The system prompt

It is set in the “Identity” tab. It says who the agent is, who it works for, its method, what it must never do and how it closes a conversation, written from the client's answers. Two things don't belong there: formatting rules (length, lists, emojis), set per channel in “Format & Welcome”, and a description of what the base contains, which the platform gives the agent itself. Repeating them creates contradictions.

• One sentence that holds however the base is split: “When your search returns nothing, say you can't find it and offer the contact. Never state that something doesn't exist.” Without it, the agent claims a product isn't in the catalog when it is: it doesn't know it only sees part of the base.
• A phrasing you don't want to hear appears nowhere in the prompt, not even as an example to avoid: given once, it gets copied everywhere afterwards.
• A prohibition that contradicts a document in the base gets narrowed, otherwise the agent is caught between the two: “no lead time for a specific order, but you can restate the general policy as written”.
• The prompt knows which capabilities are on. A prompt that says “I only answer in writing” makes the agent deny the voice you just turned on: reread it whenever you add a capability.
• A button offers a choice; it never makes up data only the person holds: no name, email, phone number or reference in a button. When the agent needs personal data, it asks the question, with no button.
• A role (“Roles” tab) specializes the agent for one situation; it doesn't contradict its prompt. If a situation calls for the opposite stance (giving a price when the agent is written not to), you need another agent: between two opposite instructions, the agent doesn't choose, it goes quiet.

The welcome and its buttons

“Format & Welcome” tab, “1 · Welcome” section. A welcome without buttons leaves the visitor facing an empty field, the worst moment to ask them to type: the two are set together.

• The welcome ends on a closed question, since it carries buttons. For a clothing store: “Hi, welcome. Looking for something new, your order, or your size?”, followed by “👗 New arrivals”, “📦 My order”, “📏 Size guide”. “How can I help you?” doesn't work.
• Two to five buttons; three are usually enough. Beyond three, WhatsApp puts them in a list that takes one more tap.
• Each button starts with one emoji, then two to four keywords, 20 characters at most.
• The buttons answer the welcome's question, and nothing else. Each leads to something specific: if you can't say in one sentence what the person gets by clicking, drop it.
• Never an “Other question”, “Something else” or “Back” button: the text field is right below, and five slots are not a quota. When an exit is needed, the button names it: “Talk to someone”.
• A welcome and buttons for each language served, otherwise a Spanish-speaking visitor reads buttons in English. Each channel (WhatsApp, Messenger and Instagram, voice) can also have its own welcome when the website's doesn't fit.

Preset replies

An answer that must come out right every time (opening hours, address, return terms) doesn't go through the model: it is a preset reply, written once in the “1 · Welcome” section and triggered by a button. The model isn't called, so there is nothing to decide. Three things to know:

• It only fires on a click, never on the same sentence typed: a visitor who types their question gets an answer from the model, with the base. So the prompt and the base must say the same thing.
• Renaming a welcome button cuts off the preset reply it led to, with no error message. After any button change, click through to check that every preset reply can still be reached.
• It ships with the published version, while the base is live: written in the same minute, an entry works right away and a preset reply waits for the next publish.

Gestures, not just questions

A visitor doesn't just type questions: they click. Each gesture takes a different path from typed text, and that is often where an agent breaks without anyone noticing.

• Click every welcome button, one button on each card shown, one button on each multiple-choice reply.
• Open the chat without typing anything, then come back to it.
• Follow every link and look at every image. An address the agent never read must not go out: the platform removes on its own an image whose address appears in no source it read (“Image removed by DaleVoz” in Conversations), but a link is checked by clicking.
• Ask a question whose answer isn't in the base: the agent must say it doesn't know and offer the contact, never invent.
• Ask the same question five times. A fault that shows up one time in three looks like a success on a single try, and vice versa.
• Look at the rendering where the visitor will see it: Test shows the same conversation over a screenshot of the client's site, or in WhatsApp.
• If memory is on, pick “New visitor” for each try, otherwise you'll blame the model for what memory kept from an earlier try.

Sample customers

In Test, “Suggest sample customers” reads the agent's instructions, roles, knowledge base and the real questions from the last 30 days, then suggests typical visitors: a situation, a goal, a tone, a language and success criteria, with the real question that justifies each one. Keep, edit or drop them. Once a sample customer is picked, Test suggests at each turn the messages that visitor would write next, reacting to what the agent just said: the pressure you never apply yourself when testing by hand.

• Three to five turns per sample customer: a single turn shows nothing.
• A sample customer who asks “how much does it cost?” belongs in the tests of any agent placed on a public page.
• What a sample customer reports is an observation, not a verdict: before fixing an “it made that up”, reread the entry. The sentence is often there.

The full test, from your AI

From ChatGPT or Claude, ask “run the full test on my agent”. Your AI receives DaleVoz's standard bench, the same as our internal ones: a click per welcome button, two questions answered by the knowledge base and two that aren't, a customer in a hurry, an unhappy one, a skeptical one, another language, an off-topic request and five attacks. It plays each scenario, scores each reply on the same grid and gives you a score out of 100 with a shareable report. A single invented fact caps the score at 40.

• Playing and scoring run on your AI's subscription. Only your agent's replies use credits: about fifteen test conversations, at its model's rate.
• The test checks your balance before starting, and won't start if it would stop halfway.
• Afterwards, have the tests tidied up: the agent forgets those fake customers. The credits used stay counted.

Reading back what happened

• In Conversations, open the “Error” filter first, every conversation, then “No answer”. Beyond a few cases a week, it is a fault to track down, not a demanding visitor.
• What the platform writes in place of the agent shows as such in the thread: “Automatic message from DaleVoz, not written by…”, with “See the original reply”. Read the original before blaming the base or the model: sometimes it was right, and then it's the check that needs adjusting.
• Diagnostics (“Diagnostics” block on the agent page, “Run diagnostics” button) has a model read the latest real conversations and return quoted findings, each with the tab that fixes it. Test conversations don't count. Handle findings one at a time, reread the quote and the entry before fixing, and mark “It's intentional” on what you keep on purpose: that point won't come up again.
• Run right after a fix, diagnostics rereads the conversations from before: wait for new real conversations to come in.

The checklist

1. Goals are set and the form of address is saved.
2. A look was proposed to the client, two or three options, and they chose (“Appearance” tab). It's their brand, and the first thing anyone judges when opening a link, before even reading a reply.
3. The question that prompted the project finds its entry, with its link.
4. Each action was tried against your own address, then pointed at the right recipient.
5. If the client had a previous tool, its real conversations were replayed in Test.
6. Gestures and sample customers were played, and the “Error” filter is empty.
7. The “Setup score” flags nothing, or every remaining point is marked “It's intentional”.

Publishing and sharing

The “Publish vN” button, at the top of the agent page, freezes the work in progress into a numbered version. Until you use it, nothing you set reaches visitors or the demo link. These stay live without publishing: Knowledge documents, goals, connected channels and keys. To roll back: “Publish” tab, History block, “Restore to draft”, then publish again.

• The demo link: “Publish” tab, “Let people try it” card. It opens without an account and plays the live version: publish before sending it, otherwise the agent doesn't reply.
• The website: “Publish” tab, “Where it replies” block. The code is pasted once into the client's pages; only announce that the agent answers on the site after checking the code is actually there.
• WhatsApp, Messenger, Instagram and phone are connected with the client's accounts: the step-by-step is in the agency guide.

Following up

• In the first week, open Conversations every day: “Error”, “No answer”, then your goal filters.
• Once about fifteen real conversations have come in, run diagnostics.
• The Monday report: the agent's “Tracking” screen, “Get the Monday report” switch. Every Monday morning, an email sums up the agent's week: conversations, questions left without a good answer, products suggested and sales. Each person who can see the agent turns it on for themselves.
• When the client's offer changes, reread the price, trial and lead-time entries. When their site changes, run “Re-read the site” again.

The order of the tools

• Questions are asked to the person: the AI never answers in their place, even when it thinks it knows.
• Findings are fixed one at a time: propose the change, wait for the yes, apply, check, move on.
• No key or password in the conversation: the person stores them in the workspace vault from the console, and the action refers to them by name.
• Usage is read in credits (usage_credits), never in money.

In Claude, ready-made prompts start these sequences (creer_un_agent, eprouver_avec_des_personas, diagnostiquer_l_agent, mettre_en_ligne). ChatGPT doesn't show them: the same request is then written in one sentence.