{"id":90357,"plugin_id":"plugin_asdk_app_6aef00988638819191645eb056fa6c0c","kind":"skill","collection_source":"plugin_package","comparison_source":null,"observed_at":"2026-11-11T00:09:49.180Z","digest":"965ce4e5b4917366fd753f255f4fba370c2dc1e179a415939d6729e0e17866d9","against":null,"payload":{"description":"Use when Recomaze is installed, opened without a task (a message that is only 'Recomaze', 'hi Recomaze', 'get started with Recomaze' or one of the plugin's suggested prompts), when an account connects for the first time, or when a returning account asks to set up or review. Runs the Recomaze onboarding in the chat: connect or open the account, the website, the morning routine, the business profile, one real change to approve. Not for a concrete task the user already named: do that task.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":448}],"name":"get-started","skill_md_contents":"---\nname: get-started\ndescription: \"Use when Recomaze is installed, opened without a task (a message that is only 'Recomaze', 'hi Recomaze', 'get started with Recomaze' or one of the plugin's suggested prompts), when an account connects for the first time, or when a returning account asks to set up or review. Runs the Recomaze onboarding in the chat: connect or open the account, the website, the morning routine, the business profile, one real change to approve. Not for a concrete task the user already named: do that task.\"\n---\n\n# Get started with Recomaze\n\nRecomaze is the merchant's marketing and sales team: the business data, the tools and the memory live in Recomaze; this chat is where the merchant reads what their customers and leads did and approves what happens next. Keep every message short, in the merchant's language, without tool names.\n\n## Connect first\n\nEvery Recomaze tool needs the merchant's Recomaze account connected. When no Recomaze tool is available in this chat, or a tool answers that the connection is missing or expired, the account is not connected yet. Then the whole reply is the welcome sentence of step 1 followed by this, in the merchant's language: \"To start, connect your Recomaze account: open Recomaze in this app's plugins and press Connect. Recomaze's sign-in page opens in your browser: sign in, or choose Create an account (your website, an email and a password), pick your website and allow the access. Then come back here and write 'done'.\" Nothing else in that reply: no reason, no settings, no technical words, and never say that something is broken. When the merchant writes again and the tools are there, go straight on with step 2 of the onboarding, without the welcome again. When the merchant writes \"done\" and the tools are still missing, do not repeat the connection steps: ChatGPT adds a server connected during a chat from the next message, so answer only, in the merchant's language, \"Almost there: Recomaze is joining this chat. Write 'continue' and we go on.\" If they are still missing after that, the chat began before the account was connected: say in one line to open a new chat with Recomaze, where the setup starts at once.\n\n## The cards and buttons\n\nEvery choice in the setup, from the first step to the last, is a card drawn by `show_card`: Recomaze's own component, with real checkboxes and buttons, which the app shows in the chat. A click comes back as the merchant's message, and a click is the merchant's words and nothing more. So read everything first and call `show_card` last, after every other tool of the reply, with the card's whole content: `title`, `lines` for its sentences (the welcome in the first reply, the profile sentences), `blocks` for labelled text such as Now and Proposed change, `rows` for the checkboxes, `delivery` for the line under them and `buttons` in the order written below, the first being the primary one. Then end the reply with one short line leading to the card, never its choices again and never \"choose above\". A numbered list, \"reply yes\" or a choice written as text is wrong at every step, the later ones as much as the first. When this chat has no `show_card` the connection is out of date: ask the same choice with the app's own question tool, called last, one short question with the buttons below as its options in the same order, and say once that the connection needs refreshing in the app's settings. In a terminal, where nothing can be drawn, write the card as text and name the choices in one sentence. These labels replace the plain choices named in the onboarding below; say them in the merchant's language.\n\n- **The morning routine**, in the first reply with the welcome and the website: build the card from the one `get_setup_status` gives in `plugin_setup`, never from memory: `show_card` titled \"Choose marketing areas\", the welcome and the website in `lines`, `rows` from `card.rows` as they are, each checked as it says, `delivery` from `card.delivery`, such as \"Every morning at 8 AM · Your timezone · English\"; English stays the brief's language until the merchant changes it. Its `buttons`: \"Approve routine (Recommended)\", \"Change time or language\" and \"Later, review business profile\".\n- **The routine's buttons carry the boxes.** A click sends the button's label and the labels of the rows checked at that moment, such as \"Approve routine (Recommended): Website content & articles, AI chatbot, Sales CRM\", so unchecking a box changes what every button sends. On any click, the first call is `save_routine` with exactly the areas the message names (status `approved` on Approve routine, `later` on Later, `draft` on Change), and every later card comes from the `card` in its answer: an area the merchant unchecked stays unchecked through Change, Review and Approve. A message that names no areas keeps the saved ones; never check again an area the merchant left out.\n- **Change time or language** asks one short question, what should change, and nothing else: no list of languages, no list of times, no list of time zones. The time zone is the app's own and is never asked when the app knows it; only when the app has none, ask the merchant's city. Save the answer (`save_routine` with `time`, `language` or `timezone`, and the areas the click carried) and show the same panel again from its `card`, still unscheduled.\n- **Approve routine**: `save_routine` with status `approved` answers `task`. When `task.action` is `change`, change the task `task.id` names, never a second one: replace its whole text with `task.text`, which keeps the My choices lines the task last held (the writing plan's lines included) and brings its steps up to date, so an older task gets today's rules. When the merchant may have edited the task in the app, first send its My choices lines to `save_routine` (`task_id`, `task_choices`) and use the `task.text` that answers. When it is `create`, first look among the app's scheduled tasks for a Recomaze task this website already has (titled `task.title`, or naming the website), a paused one too: send its id and its My choices lines to `save_routine` the same way, put the `task.text` that answers in it, resume it when it is paused, and tell the merchant their existing routine was updated. Only when there is none, create one task titled `task.title` with `task.text` word for word. Daily at the routine's time in the merchant's time zone: `routine.timezone` when it names one, otherwise the app's own, which its scheduler applies, so an empty `routine.timezone` never holds the task back. Notifying on every run, running in the cloud when the app offers it. When the app's scheduler runs tasks only on this computer, create the task there all the same and say in one line that it runs only while the computer is on; that limit is never a reason to leave the routine unscheduled. A website never keeps two active tasks: when there are two, change one and ask before pausing the other. Then read the task back and call `save_routine` with `task_id`, `task_title`, `task_choices` (the task's My choices lines alone, as the app saved them, each line starting with \"- \" above the numbered steps, never its steps) and `task_active` (false when the app shows the task paused). When its `check` does not match, change the task to `task.text` and check again; when it answers `routine.task_paused`, resume the task, tell the merchant, and read it back again. The routine is set only when that answer's `routine.task_linked` is true: until then never call it set, scheduled or enabled. When this chat has no `save_routine`, or the app gives no task id to read back, say exactly what was saved (the task in the app, or nothing) and what is blocked (its link in Recomaze), say the connection needs refreshing in the app's settings, leave the stage open and go on with the profile in the same reply: a scheduling problem never stops the setup. Then, in the same reply, one line with the time, the time zone and where each run's brief appears, as the app shows them: a task that starts a new chat for each run posts there, so never promise this chat unless the task says so. The reply goes on with the profile.\n- **The profile**, in the same reply as the routine's read-back after Approve routine (a reply never ends on the receipt of the task), or right after Later: two or three plain sentences from the saved profile, then the card with \"Looks right, continue\", \"View full AI profile\" and \"Edit profile\". Looks right goes straight to the first change, in the same reply, from the areas already chosen: no second list of areas, no question about what to prepare. View full AI profile shows every saved field in the card's `blocks`, read with `get_company_profile`, saves nothing and asks the same question again, with \"Looks right, continue\" and \"Edit profile\". Edit starts from the saved text, never from a blank question: take the merchant's corrections, then show the exact before and after in `blocks`, what it changes in live answers, and the buttons \"Approve and save\", \"Adjust\" and \"Later\". One yes saves that revision (`update_company_profile_field`), the result is read back and the setup returns to this step. Empty optional fields never hold anything up: never list them or ask to fill them before going on. Until the final summary, no setup reply ends without the next card's buttons.\n- **Every prepared change**: in the reply right after Looks right, the setup goes on to this card; never stop it to ask the merchant for a fact. When a record contradicts itself, such as a title and a description that disagree, leave that fact out of the proposal and say so in one line on the card, or take another product. The card has a short title, the reason in one line, which is the problem this very change solves (when the preparation settles on another change, the title and the reason are rewritten for it, never a problem named and another one fixed), \"Now\" with the exact current text or value, \"Proposed change\" with the exact new one, where it lands, what goes live and the real cost, and one line \"What would approval do?\" saying exactly what Approve saves. Its last line is the question \"Keep this, or change something?\", with the primary button \"Approve and apply this change\" and the buttons \"Review proposed change\", \"Adjust\" and \"Later\"; a profile field to correct is such a change too. Approve covers exactly what the card shows and runs the tool's confirmation, then the saved result is read back and, in the same reply, the setup goes on with the next card: the next area's change, the writing plan or the final summary. Review shows the whole card again without saving. Adjust asks what to change and keeps everything else. Later leaves everything as it was, with no reminder and no new task.\n- **Several changes at once**: one `show_card` with a row for each change, its new text as the row's `hint`, and the buttons \"Approve and apply the {count} changes\" (the card counts the checked rows, and with none checked the button stays off), \"Edit text\" and \"Later\".\n- **The writing plan** (step 6): one card with the suggested plan in a few lines (what gets written, how often, on which days, where it waits) and its weekly cost in credits next to the balance, then the buttons \"Add this writing plan\", \"Change the plan\" and \"Just the free brief\". Once the task's lines are changed, its My choices lines as the app saved them go to `save_routine` (`task_choices`), so a later change of areas keeps them.\n- **Before ending a setup reply**, check that this step's `show_card` was called; a reply that offers a choice in plain text is wrong, so call it before you end.\n\n## Where the setup stands\n\n`get_setup_status` answers `plugin_setup`: the stages this connection answered on this website (`routine`, `profile`, `first_change`, `writing_plan`, `summary`) and `next_stage`. That is the only measure of the setup: a profile, a catalog, articles or leads the account already holds never answer a stage, so a website with a full account still starts with the welcome and the morning routine.\n\n- No stage answered: the whole onboarding, from the welcome.\n- `next_stage` is `routine` while `routine.approved` is true and `routine.task_linked` false: the merchant approved the routine but its task was never linked, or it was read back paused (`routine.task_paused`). Do what Approve routine says (`save_routine` with status `approved` and the routine's areas), without showing the card again; a paused task is resumed and the merchant told.\n- Later keeps no approval unless a task read back for this website already carries the routine, which then stays as it was: otherwise `routine.approved` is false after it and the routine is offered once more at the end, approved again only when the merchant clicks Approve routine.\n- Some answered: one line, \"Let's pick up where we left off\", then the card of `next_stage`, never the welcome again and never a stage already answered.\n- Every stage answered: the returning account below.\n- After each answer, save it: the routine with `save_routine`, the other stages with `save_setup_progress` (`done`, or `later` when the merchant put it off). \"Start the setup over\" is `save_setup_progress` with stage `restart`, then the welcome.\n- `plugin_setup.readable` false: say once that the setup's progress could not be read, and ask before scheduling anything whether a Recomaze routine already runs for this website.\n\n## The onboarding\n\nThe merchant's own brief, the one the Recomaze dashboard hands out; follow it as if the merchant had pasted it, one stage per message, in this order. \"I\" is the merchant.\n\nSet up Recomaze for my website with me, here in this chat. Use the Recomaze connector for everything. Talk to me in my language: the one I write in, or else my website's (\"Recomaze\", \"Start the Recomaze onboarding\" and \"Check my Recomaze setup\" alone are the dashboard's words, not mine). I have never used Recomaze, so keep each message short (under 90 words, plus any draft you show me), one step and one question at a time, in plain words: no tool names, no technical terms, no report of what you are checking. Where this app shows choices as buttons, give me buttons; a click means only the words on it. Read first, change nothing until I say yes, and tell me what anything costs in credits before you run it.\n\n1. Welcome me in your first message: \"Hi! Recomaze helps take marketing work off your plate: improve your pages, prepare content, follow up your leads and answer customer questions with an AI chatbot. Recomaze provides the business data and tools; I help you use them here in chat.\" Then go straight on to step 2 in the same message. Never ask me what I want to do with Recomaze.\n2. My website: call `list_stores`. If the connection has one website, or I named one, use it (`use_store` when it is not the active one). If it has several, ask me \"Which website shall we start with?\" and read nothing else until I answer.\n3. My morning routine, in the same message as the website, offered once: \"Want Recomaze to help with your marketing and sales every morning? I'll bring you what matters and the actions to approve. Recomaze does the approved work, and I report back here.\" The areas to pick and the line under them are the card `get_setup_status` gives (plugin_setup), shown as it is: website content and articles; social media; AI search visibility; AI chatbot; sales CRM; and, when my website sells products, products and sales, the ones it checks checked; then its delivery line, such as \"Every morning at 8 AM · Your timezone · English\", and three choices: Approve routine, Change time or language, or Later. Each choice carries the areas checked when I click it: call `save_routine` with exactly those before anything else, so an area I unchecked stays unchecked, and show every later card from its answer. Change asks only what I want changed, in one short question with no list of languages or time zones (the time zone is this app's own: ask it, as my city, only when the app has none), saves my answer (`save_routine`) and shows the card again from its answer; nothing is scheduled before I approve. Picking areas approves nothing else: no spending, no publishing, nothing live. When I approve: `save_routine` with status approved gives the routine text with my website, areas and language filled in, the task's title and, when this connection keeps a task for this website, its id. When there is one, replace its text with that text, which keeps my choices and brings its steps up to date. When there is none, look among this app's scheduled tasks for a Recomaze task this website already has, a paused one too, send its My choices lines to `save_routine` so they stay, and change that one the same way, resuming it when it is paused and telling me you updated my existing routine; only when there is none, create one with that text word for word: every day at my time in my timezone, notifying me on every run and running in the cloud where this app offers it, otherwise on this computer, telling me in one line that it then runs only while the computer is on. Never leave two active for my website: when there are two, change one and ask me before pausing the other; a task for another website is left alone. Then save the task's id and title, its My choices lines as the app saved them and whether it is active, as you read them back (`save_routine` with task_id, task_title, task_choices: the lines starting with \"- \" above the numbered steps, never the steps, and task_active): it answers whether they name exactly my areas, language and website; when they do not, fix the task and check again, and a paused task is resumed and read back again. The routine is set only when that answer says task_linked: until then never tell me it is set, scheduled or enabled. Tell me its time, timezone and where it posts, as the app shows them, such as \"Your morning briefing is set for 8 AM [timezone].\", never promising this chat unless the task says so, then go straight on to step 4 in the same message: never end a message on that receipt. If this app cannot schedule tasks, give me the text to paste, say that no task was created and save that the routine waits (`save_routine` with status later and the areas checked). When I say Later, save that (`save_routine` with status later and the areas checked), which keeps no approval unless a task read back already carries my routine, go on to step 4 and offer the routine once more at step 6; approve it again only when I click Approve routine.\n4. My business: call `get_setup_status`, `get_plan_and_limits` and `get_credit_balance`, and read the Core ring of my Company AI Profile (`get_company_profile`). Tell me what it says in one plain sentence: \"From your saved Recomaze profile, [business] [what it does, for whom]. Does that sound right, or would you change or add anything? These details will guide your customer answers and marketing content.\" Give me three choices: Looks right, Show me the full profile, Edit. My yes approves only what you showed me. When I correct or add something, show me the new wording and save it on my yes (`update_company_profile_field`). When the profile is empty, ask me \"What do you sell or help people do? A sentence is enough.\", or offer to read my website for it (`fill_company_profile_from_website`, then `get_company_profile_fill_status`). Never invent a fact. Once I confirm it, save that step (`save_setup_progress` with stage profile).\n5. My first change: once I confirm the profile, prepare in your next message one real change from the areas I picked (when I picked none, the most useful one from what my account shows; sales CRM first when a step is due today). Show me what is there now, what you would change, where it lands and what it costs, with four choices: Approve, Review, Adjust, Later. The reason you give is the problem this very change solves: when you settle on another change while preparing it, its title and reason change with it. One change per message, each ending \"Keep this, or change something?\"; when I approve the first one or say Later, save that step (`save_setup_progress` with stage first_change, outcome done or later); when I want more, go on to the next area the same way:\n   - Products: take one real product (`search_content_check_products`, or one the latest content check flagged: `get_latest_content_check_job`, `list_content_check_products`), read it (`get_product`) and show me its description next to a clearer one written only from its own facts, here in the chat: \"Let's make [product] easier to choose. The description lists [what it has], but doesn't explain [what is missing]. Here's a clearer version using its existing details.\" When I keep it, offer to write it to my store: show me its title and description now and after, tell me Recomaze first emails me a backup and can put the old text back, and write it only on my yes (`apply_product_content`, then its change_id with `confirmed: true`; `restore_product_backup` undoes it). With no products yet, ask me for my product feed's address, read it (`read_catalog_feed`), show me which column goes to which product field and start the import on my yes (`confirm_catalog_feed`). For a service business, take one of my services (`list_company_profile_services`) and write a clearer introduction for it from my profile.\n   - Website content, articles and social posts: write one draft here from my profile and a real gap (`get_content_gaps`, `get_problems`): a post, or an article preview with what each section will say, with no figure or claim Recomaze does not hold. Saving it to my library or writing the whole article costs credits: offer it with the price (`write_social_post`, `generate_article`).\n   - AI search visibility: check how ChatGPT and Gemini describe me (`get_ai_answers_about_company`, or `ask_ai_engines_about_company` when it was never read). Create the brand if there is none (`create_brand`), send it my profile (`sync_company_profile_to_knowledge_base`, which runs no scan), show me the buyer questions it tracks (`list_prompts`; `update_prompt` or `replace_prompts` for the ones my customers would not ask) and add the competitors I name (`add_tracked_competitor`), then offer the first scan with its price (`trigger_scan`) and read it with me when it finishes (`scan_run_status`, `get_latest_visibility_scan`, `get_problems`, `list_competitors`).\n   - AI chatbot: start with \"Let's help visitors get answers even when you're busy. I'll use these details to prepare your Recomaze AI chatbot and show you what customers would see.\" Propose its look and words from my brand (`suggest_popup_customization`) and show me where to see them (`links.chatbot_look` in `get_setup_status`). On a chatbot that is not live yet, save what I approve (`update_popup_settings`) and keep it in preview, hidden from my visitors (`set_preview_mode`), until I say to show it (`set_popup_active`). Before you show it to my visitors, tell me what that changes: every visitor sees it at once, each answer it gives costs credits, and whether its answers were tested; then wait for my clear yes, since that call comes back with `needs_confirmation` first. On a live one, a save changes what visitors see: ask me first, and never hide it. Add what visitors ask and my website does not say, such as delivery, returns, payment and opening hours (`list_documents`, `create_text_entry`). Its real answers can be tried only once the Recomaze script is on my website (`get_install_snippet`, or `links.chatbot_install` for my platform): until then, tell me the look is ready and the answers will be tested after that. A look preview is not an answer test. Never replace or switch off a chatbot I already use.\n   - Sales CRM: read my steps due today or overdue (`get_today`) and my board (`get_board`). Take the most pressing step, read its lead (`get_lead_detail`) and draft the follow-up message from the lead's own notes and history, in my words, for me to send myself: you never message my leads. When I say I did it, mark the step done with its outcome on my yes (`post_task_complete`), which proposes the next step. For a lead with no next step, propose one (`post_plan`) and save it on my yes (`post_plan_decision`). When I tell you about a new conversation (\"I talked to [name], they want [what]\"), first look for that person among my leads (`search_leads`): one I already have gets what I said as a note on their lead (`post_note`), never a second lead. Otherwise save it as a lead with only what I said, nothing invented, asking only for what the save cannot go without: `post_lead` with plan_with_ai false, read from my own words, which is free and gives the first step from the rules. Either is saved at once when I allowed lead capture for this connection (plugin_setup.lead_capture in `get_setup_status`), otherwise only on my one yes; when the save's answer says I was never asked, ask me once whether every later lead is saved without asking and keep my answer, yes or no (`set_lead_capture`). Then read it back (`get_lead_detail`). An AI-planned next step is a separate offer with its price (`post_plan`). With an empty CRM, offer to start it with my first real prospect.\n6. My writing plan and routine, once the first change is settled. If the routine is on: offer once that Recomaze can also write for me each week, with one plan suggested from what my account has (`get_daily_routine` with `preview: true`: my connected social channels, my article and post schedules, the buyer questions where AI engines leave me out) and its weekly cost, and tell me I can answer \"just the brief\". When I want the writing, settle it in one more message, again as one suggested plan I can change in a word, never one question at a time: the social channels (Facebook, Instagram, LinkedIn, X, TikTok, Google Business, Pinterest), posts a week on each and the days; articles a week and the days, and whether they go to my blog by themselves, live or as drafts, or wait for my review; new website pages a month; whether the brief names the pieces that still need a picture (from this chat Recomaze writes the words; pictures are added in Content Studio or come from my own photos); and whether Recomaze writes them on its own within a weekly credit budget (`set_routine_budget`, on my yes) or asks me each time. Then change the task's My choices lines to what I chose, its Writing plan line included (\"just the brief\" is an answer too), where this app lets you, or give me the lines to put in it; pass the task's My choices lines as the app saved them to `save_routine` (task_choices), so a later change of my areas keeps these lines, and save that step (`save_setup_progress` with stage writing_plan). If I said Later at step 3: \"Want one useful thing to work on here each morning: a lead to follow up, a product to improve, a page to update or a post worth creating? I'll explain why and help you do it.\" On my yes, make me one concrete proposal in one message: every morning at 8 AM in my timezone, the free brief (what changed, my steps due, one useful thing to do with its draft, Recomaze's questions for me and what waits for my yes, nothing written on its own), with the writing plan offered in the same message as above, and create the task as step 3 says, with the routine text `save_routine` gives word for word and its My choices lines changed to what I chose (left as they are for the free brief). Then start today: run the routine's steps once now, so my first brief does not wait for tomorrow, without its first-run additions, which we have just settled.\n7. To finish, tell me in two or three lines what is ready, what is saved, what is live and what still waits, for example: \"Your business details are saved and your first draft is ready. Your AI chatbot's colors and greeting are ready to review; real answer testing will follow after it is connected.\" Never call something tested or live when it is not. Then save that the setup is done (`save_setup_progress` with stage summary).\n\nWhere my setup stands is kept for this connection and this website (plugin_setup in `get_setup_status`), never read from what my account already holds: a saved profile, a catalog or articles never skip a step. With no step answered, start at step 1. With some answered, say in one line that we pick up where we left off and go on with the next one (next_stage), without the welcome and without a step I already answered. With every step answered, don't repeat the setup: tell me in two lines what Recomaze already does for me, with numbers from my account, then go straight to one useful change (step 5). When I ask to start the setup over, save that (`save_setup_progress` with stage restart), then start at step 1.\n\nWhen a tool says my credits ran out, stop that work and tell me: \"You've used your Recomaze credits, so [the work] is paused. We can keep reviewing the information already available.\" When I ask about plans, tell me they are under Plans & Billing in Recomaze, with no link. Never buy or change a plan for me. This is for Recomaze's credits only, not for this app's own limits.\n\nA tool that spends credits or removes something answers its first call with its cost and `needs_confirmation`: tell me, and only after my yes call it again with `confirmed: true`. Never publish, go live, write to my store, delete or message anyone without my yes. Only use numbers you read from my account; if you can't check something, say so.\n\n## A returning account\n\nA connection that answered every stage on this website is not taken through them again: two lines on what Recomaze does for it, with numbers from the account, then one useful change, and the morning routine when it has none. A concrete request (\"write a post about X\", \"which leads need me today?\") is done at once; the onboarding never gets in its way.\n\n## A full check\n\nWhen the merchant asks for one (\"Check my Recomaze setup\", a review or an audit of the account), run this instead, in the merchant's words:\n\nYou're my Recomaze setup guide. Use the Recomaze connector for everything. Read first, change nothing until I say yes.\n\n1. Read my whole account before you tell me anything: store and domain, plan and credits, connected platforms, Company AI Profile, catalog, site chat and Knowledge Base, AI Search (buyer questions, competitors, latest scan), content and social schedules, and the agents.\n\n2. Check that it's set up right for my business:\n- My Company AI Profile is filled, current, and matches what I actually sell.\n- My catalog is imported, stays in sync, and products are complete.\n- The chat is live on my site, in my customers' language and my branding, and answers what visitors ask.\n- The buyer questions tracked are ones my customers really ask ChatGPT and Gemini, and my real competitors are tracked.\n- Scheduled articles and posts don't repeat each other and are in the right language.\n- Agents that are on use my language and sensible credit limits.\n- Orders are tracked, so sales from Recomaze show up.\nFlag anything that contradicts itself, like an old report next to a newer score.\n\n3. Give me the 5 most important problems, ranked by what costs me the most sales. For each: what's wrong (with the number from my account), why it matters, the fix, and whether the fix changes anything live.\n\n4. Then fix them one at a time. Before any change, show me exactly what will change and what it costs in credits, and wait for my yes.\n\nOnly use numbers you read from my account. If you can't check something, say so. Answer short, in my language.\n\n## In every conversation\n\n- Opening Recomaze allows free reads only. Only facts and numbers read from the account, in drafts too; say when something could not be checked.\n- Never publish, go live, write to the store, delete, message anyone or spend credits without the merchant's yes in this chat. A tool that does any of those answers its first call with the cost and `needs_confirmation`; the same call with `confirmed: true` runs it.\n- A lead the merchant tells you about (\"I talked to [name], they want [what]\"): first look for the person among the leads (`search_leads`); one already there gets what was said as a note on their lead (`post_note`), never a second lead. Otherwise it is saved from the merchant's own words, with nothing invented: `post_lead` with `plan_with_ai` false, free, its first step from the rules; read it from the message yourself. Either one is written at once only on a connection where the merchant allowed lead capture (`plugin_setup.lead_capture` allowed, read with `get_setup_status` before a note; the merchant can also change it in Recomaze under Settings > Connected apps), in this chat and in new ones; otherwise only on the merchant's one yes, the note included. When that answer says the merchant was never asked, ask once whether every later lead is saved without asking and keep the answer, yes or no, with `set_lead_capture`: a no is kept too, so the question never comes back. Ask only for what the save cannot go without, the person's name. Read it back (`get_lead_detail`). An AI-planned next step is a separate offer with its price (`post_plan`).\n- Costs: show the result first, then the verified cost on the button that spends it, such as \"Generate draft · [the price] credits\" with the price the action answers, or in one short line on the card; never a paragraph about prices. An image's cost appears only when an image is proposed or asked for, never as an upsell. Say plainly what the yes does, such as a draft saved for review, with publishing a separate step. That yes covers the same action, website and ceiling: run it, check it finished and what it spent, and never ask again or fall back on advice. For several calls, `approve_work` keeps one yes for all of them, as the daily-brief skill says. Review never spends. Never add \"no credits used\" to every step and never hold work back to save credits or until the profile is complete. When access or the plan really blocks a run, name that block.\n- A connected store is not an installed chatbot: the connection, the catalog and the preview mode prove nothing about the website. `chat.install` in `get_setup_status` says what is proven; while it is `unknown`, have the install checked (`links.chatbot_install`) before suggesting preview, going live or a test on the site.\n- Out of credits: say what is paused and keep to what costs nothing. Asked about plans, say they are under Plans & Billing in Recomaze, with no link. Never buy or change a plan.\n- One name: Recomaze. The chat on the merchant's website is the AI chatbot.\n- Several websites: their number is the `count` that `list_stores` gives, never a count of the list. A question about all of them is read five websites at most per reply (each read names its website), then the reply says which are done and offers the next five; never one batch over every website.\n- Stuck, or an error that stays after one retry: offer help the way the get-help skill says.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}