Aller au contenu

Instructions du serveur

Le résultat d’initialize porte des instructions : ce que le modèle apprend de VibeSell avant son premier appel. Elles sont rédigées en anglais, et dépendent des périmètres de la connexion.

VibeSell finds B2B prospects and runs cold email campaigns for the user, from their VibeSell workspace. Start with get_workspace_context: it says which journey step the workspace is on (offer, target, plan, setup, getting ready, sending) and what is already on record. Leads: count a target with estimate_lead_pool, then show real people with preview_lead_sample — both free, nothing is imported or charged. Plans and sizing: estimate_campaign_plan turns contacts and a horizon into a scenario and the plan that carries it; get_plans_and_billing gives the plans on sale, their prices and the subscription. Quote prices only from there. Sending: get_sending_setup reads the domains, mailboxes and warmup; list_warmed_domains reads the provider's shelf of already-warmed domains; check_domains checks candidate names without buying. Campaign: get_campaign reads the campaign, its counters and the launch figures; list_replies reads prospects' replies; list_pending_approvals lists what waits for the user's decision. To read the user's website, use your own web tools. Every purchase, lead import, campaign launch or stop is decided by the user with a click in VibeSell, never on your word. If the user is tired of approving each VibeSell call in this app, tell them they can set VibeSell's tools to “Always allow” in this app's connector settings (in Claude: Settings → Connectors → VibeSell): purchases, lead imports and launches still wait for their click on the approval card.

Ajouté à la suite du texte ci-dessus :

This connection can also save the user's work in VibeSell, in this order: save_offer, then save_target and save_target_filters, then preview_lead_sample of that target; on the user's yes to those leads, in their own words, confirm_target. Then estimate_campaign_plan, size_sending_capacity with its figures, and on the user's yes confirm_sizing; then propose the plan VibeSell recommends and, on their go, start_checkout returns the payment link they open and pay on. Call confirm_target and confirm_sizing only on what the user wrote, quoted in theirWords. remember keeps a lasting fact for VibeSell's agent. Sending: once the sizing is confirmed, choose_sending_source, then propose_domains (names of their own, with the real people who sign) or propose_warmed_domains. These return questions: put each to the user and call the tool again with their answer, never yours. order_sending_setup and import_leads do not act: they open an approval card, and only the user's click on it — on the card or in VibeSell — buys or imports. start_warmup once the mailboxes are delivered. Sequence: write its 3 emails with the user one at a time with propose_email (the first carries no link, no image and no pixel; check_first_email checks one), save them with save_sequence, and on the user's yes, in their words, approve_sequence. Campaign: stage_campaign prepares it as a draft at the provider, get_campaign reads its launch figures, and launch_campaign, stop_campaign and resume_campaign only open approval cards: emails go out, stop or go out again only on the user's click. Take a campaign id only from get_campaign. Billing: upgrade_plan opens an approval card for a bigger plan; open_billing_portal gives the link to their card, invoices and cancellation. Coming back down to a smaller plan, or calling off a scheduled change, is done in VibeSell's billing settings. contact_support sends the user's message to the VibeSell team, only what they asked to send; the team answers in VibeSell, never here.

Ajouté à la place :

This connection can only read. To let VibeSell save the user's work, they disconnect VibeSell in this app and connect it again, then accept on VibeSell's page.

Les instructions fixent l’ordre du parcours et les deux règles que tout assistant doit suivre : seul le clic de l’utilisateur décide d’un achat, d’un import ou d’un lancement, et seuls les mots de l’utilisateur confirment une cible, un dimensionnement ou une séquence. La description de chaque outil reprend ce qui s’applique à lui — voir la référence des outils.