Update to Recomaze

{
  "description": "Use for the Recomaze daily routine: a scheduled morning run, 'my brief', 'what matters today', 'what did my customers do yesterday', 'which leads need me today', or setting the routine up and changing its areas, time or writing plan. Reads what changed on the website since the last run and answers with one short brief: the figures, the sales steps due, a batch of useful work with its texts ready, what waits for a yes.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 302
    }
  ],
  "name": "daily-brief",
  "skill_md_contents": "---\nname: daily-brief\ndescription: \"Use for the Recomaze daily routine: a scheduled morning run, 'my brief', 'what matters today', 'what did my customers do yesterday', 'which leads need me today', or setting the routine up and changing its areas, time or writing plan. Reads what changed on the website since the last run and answers with one short brief: the figures, the sales steps due, a batch of useful work with its texts ready, what waits for a yes.\"\n---\n\n# The Recomaze daily brief\n\nThe brief is why the merchant comes back every morning: what their customers and leads did yesterday, and what to approve today, in 150 words they read on their phone. It is one message, on every run, even when nothing changed.\n\n## The routine text\n\nThe text of the scheduled task, word for word. A scheduled run sees only its own text, so it says everything a run needs; the \"My choices\" lines are the merchant's settings and change only on their yes.\n\n```\nRun my daily Recomaze routine. Use the Recomaze connector for everything.\n\nMy choices. Edit these lines whenever you want: the routine reads them on every run. As they are, every run is a free daily brief: what changed, a batch of useful work with its texts ready, my questions and what waits for my yes, and Recomaze writes nothing on its own. The posts, articles and pages I ask for below are written into my Recomaze library for me to read, and nothing is published. Each one waits for my yes, unless I let Recomaze do that kind of work on its own within my weekly credit budget (Settings > Connected apps in Recomaze).\n- Website: every website of this connection, one brief each, five at most in one run, with the others named in one line. Write one website's address to run the routine for that one only, or the addresses of the five I want.\n- Areas: website content and articles, social media, AI search visibility, AI chatbot, sales CRM, products and sales. Remove the ones I don't want: the brief covers only my areas, with what waits for my yes and my credits.\n- Language: the website's language. Write another one here, such as English, to have every brief in it.\n- Social channels: the ones connected to Recomaze. Recomaze writes for Facebook, Instagram, LinkedIn, X, TikTok, Google Business and Pinterest.\n- Posts: none. To have some, write how many a week on each channel and on which days, such as \"Facebook: 2 a week, Monday and Thursday\".\n- Articles: none. To have some, write how many a week and on which days, for the buyer questions where AI engines leave me out.\n- Blog: finished articles wait in my library for my review. Write \"live\" or \"drafts\" to have them go to my blog by themselves once my blog is connected to Recomaze: that is set once, on my yes, as Recomaze's own article schedule (create_article_schedule with its publishing choice); the daily run itself never publishes.\n- Website pages: none. To have some, write how many a month (service, category, comparison or FAQ pages) and on which day.\n- Images: name the pieces that still need a picture. From this chat Recomaze writes the words; pictures are added in Recomaze's Content Studio or come from my own photos.\n- New ideas: offer me something new when Recomaze found something worth it; on Monday, also ask me about anything else Recomaze could do for me that these lines don't cover.\n- Writing plan: not asked yet. Once I have answered, in the setup or after a brief, this line says what I chose and no brief asks again.\n\n1. Call list_stores. Run the routine for the website my choices name, or for each website in turn (use_store), five at most, with one brief per website. If the website my choices name is not on the list, say so in this chat and read no other website instead.\n2. Call get_daily_routine, once for each website, with areas set to the keys of the areas on my Areas line, whatever language I wrote it in: content (website content and articles), social, ai_search, chatbot, sales_crm, products (products and sales); with none given, Recomaze reads the areas I chose in the setup. It returns today's sections for those areas, what changed since the last run, the website's language, whether this is the routine's first run there (first_run), the tasks of each section and today, the things the brief can lead with, in their order of worth.\n3. Work through the sections it returns, in order; an area I removed is not among them, so never bring it in from anywhere else:\n   - Data gaps: the questions visitors asked that the chat could not answer, and the data missing from my Company AI Profile or catalog; the ones that matter most become today's questions for me.\n   - Conversations: yesterday's chat figures, and the conversations that matter: a buyer who left without an answer, a complaint, a competitor named.\n   - Sales: the steps due today or overdue, the leads waiting for a next step and my won and lost counts; for the most pressing step, the follow-up message drafted from the lead's own notes (get_lead_detail), which you never send yourself; for a lead with no next step, one proposed (post_plan) on my yes.\n   - Articles and website pages: what is ready, held or scheduled, the buyer questions where AI engines don't mention me yet, and, on the days my choices name, the next article or page to write.\n   - Social media: the posts ready in my library and, on the days my choices name, the posts they ask for on each of my social channels: an idea for that channel (suggest_social_posts with its platform) and the post written from it into my library (write_social_post with that platform). Name my channels that are not connected to Recomaze yet, and, when my choices ask for it, the pieces that still need a picture.\n   - Approvals: what Recomaze's agents propose and what waits for my decision.\n   - Today's theme, when there is one: AI Search on Tuesday, catalog or service pages on Wednesday, the week in review on Friday, credits on the 1st of the month.\n   - Credits: my balance and what the routine spent this week.\n4. Each task has a mode. auto: I allowed that kind of work within my weekly budget, which is my yes given in advance, so run it now with the tool and arguments it names, without confirmed, and the posts, articles and pages only as my choices ask for them today; Recomaze runs it while the budget lasts, and one that still answers needs_confirmation joins what waits for my yes. ask: do not run it; list it for me. confirmed: true is sent only after my yes in this chat, never because a task is auto. Never create another scheduled task, and never publish, delete or change anything live without that yes. This scheduled task stays on: a run finishes today's brief, never the task, so never pause, disable, reschedule or delete it, whatever the run found, unless I ask for that in this chat.\n5. Answer on every run, in the language my Language line names (the website's when it names none) unless I asked this chat for another, with a brief that starts with \"Recomaze • [business] · [today's date]\". I read it on my phone: 150 words at most besides the prepared work, which is never cut to fit. Before you answer, give check_brief the brief without that work: when it answers within_limits false, leave out the least useful lines, never a prepared text, and check again. One line per point, no explanation of your choices, and leave out a part with nothing in it:\n   - What matters today: from my areas and what could really be read, in two to four lines: what changed since the last run, my visitors' questions, objections or complaints, sales opportunities and steps due or stuck, material ready or unfinished, AI search signals and, on a store, verified catalog or product problems. Each fact must change a priority or a decision, and not every area speaks every day. Keep volumes apart from commercial results, give the date and source when freshness matters, never show old data as new, and never read missing access or data as zero. On the first run this is the starting point, not differences. My profile (offer, audience, voice, limits, goal when known) guides what you choose and write; never recite it.\n   - Done by Recomaze: what you finished on your own or on my earlier yes, each one checked, with what it cost; first finish what I already approved, within that approval.\n   - Ready for you: the interventions worth it today, chosen from today_batches and the sections by evidence, urgency, expected effect and what moves something already started. The order of the list is not a priority, and sales never comes before marketing by rule. One verified need can tie together an answer for a prospect, a fix on a page, product or FAQ and an existing piece adapted for a real channel; several independent interventions can be justified on the same day, with no quota and nothing written to fill the report. Each one: its objective, why now with the date of the data, and every target's complete text and destination, ready to use, written only from what Recomaze knows about my business, reusing what my account holds, after reading what each target names to read first; never offered for later and never cut to fit. A record that looks like a test, a sample or a demo never leads. On a store, a target is often one real product: read it (get_product) and write only from its own facts. Show each intervention on one card with show_card: a row for each target with its text as the hint, and Approve, Review and Later, the verified cost on the button that spends it. Before the card, call approve_work with the exact calls Approve would make and the ceiling; on my click call it again with confirmed true and run them without asking again. When I unchecked rows, my click is my yes for the rows left: call approve_work for those calls alone, then again with confirmed true. approve_work takes only apply_product_content, create_text_entry, generate_article, post_lead, post_plan, post_task_complete, write_social_post; any other call on a card, such as an article edit, never goes in it and keeps its own yes. A tool that takes confirmed is called before the card without it, which answers with its cost and runs nothing, and again with confirmed true on my click; a tool that takes no confirmed is called once, only after my click, with exactly the arguments the card showed. Read each result back and say what finished and what failed. Review shows the texts and the facts behind them and spends nothing; Later leaves them as they are. Publishing and messages to my customers are never part of this yes. What already waits for my yes goes on the same kind of card, grouped by channel with its cost: at most three, the rest counted in one line. Never a numbered list, and never write a button as text.\n   - Blocked or missing: what could not go on and why, and up to three short questions, one line each, about the details Recomaze is missing, from the data gaps, each with why it matters in a few words. When I answer in this chat, show me what you will save and save it on my yes, for free: an answer for visitors as a knowledge base entry (create_text_entry), a fact about my business in the profile (update_company_profile_field). Never fill a gap with a guess.\n   - Moved since last time: what advanced from last_brief, in one or two lines; none on the first run.\n   Never offer again what last_brief shows offered or put off unless something new justifies it: take the same work one step further instead. End every run with save_brief: each item this brief offered or finished and where it stands, with its id as key and where it lands as target, under the key or title last_brief gives it when it is the same work; what it does not name stays as it was, so the next run continues it.\n   Never skip the brief, and never ask setup questions in its place: when nothing changed, take an earlier idea one step further instead of repeating yesterday's. A section that could not be read is not zero: name it. Never invent figures, prices, policies or promises.\n6. When get_daily_routine answers first_run: true, end that brief with two short additions, 60 words at most together, the second only while my Writing plan line says not asked yet. First, how to get it every morning in this app: the task repeats daily at my time and timezone, notifies me on every run (on my phone and by email where the app offers it) and runs in the cloud where the app offers it; where it runs only on this computer, tell me in one line that it runs only while the computer is on; say where each run's brief appears as this app shows it, and never promise this same chat when the app may start a new one. Then one question: whether I want only this free brief, or also Recomaze to write for me each week, with one plan suggested from what my account has (my connected channels, my schedules, the buyer questions where AI engines leave me out) and what it costs a week. When I answer, change this task's My choices lines yourself where this app lets you, the Writing plan line included, or give me the lines to put in it; until then every run stays the free brief.\n```\n\n## A brief on demand\n\n\"What matters today?\" or \"my brief\" in a chat runs the same steps, with `get_daily_routine` called with `preview: true`, which does not count as a run, and answers with the same report: what matters today, then the interventions chosen from `today_batches` and the sections by evidence, urgency, expected effect and what moves something already started, each with its texts written out and its buttons, its length checked with `check_brief`.\n\n## Setting the routine up\n\nA request to set the routine up, or to change its areas, time, language or writing plan, uses the routine stage of the onboarding (the get-started skill): the card `save_routine` gives, its delivery line such as \"Every morning at 8 AM · Your timezone · English\", Approve routine, Change time or language, or Later, each button carrying the areas checked when it is clicked. One scheduled task per website: `save_routine` names the one this connection keeps, which is changed rather than a second one added (its whole text is replaced, keeping its My choices lines), a task for another website is left alone, and the task's My choices lines are read back and checked with `save_routine` (`task_choices`, never the numbered steps, and `task_active`); the routine is set only once that answer says `routine.task_linked`, never while its task is paused. Setup never replaces today's brief, and today's brief never asks setup questions.\n\n## In a scheduled run\n\n- The run completes today's brief, never the task: it never pauses, disables, reschedules or deletes the scheduled task, and a successful or empty brief is no reason to stop it.\n- The brief covers only the areas on the routine's Areas line: their keys go to `get_daily_routine` as `areas` (content, social, ai_search, chatbot, sales_crm, products), whatever language the line is written in, and nothing from an area left out is added from elsewhere; called without them, it reads the areas the merchant approved in the setup. It is written in the language of the Language line, the website's when the line names none.\n- Tasks marked `auto` are called without `confirmed`: the merchant allowed that kind of work within a weekly budget, and Recomaze runs it while the budget lasts; one that still answers `needs_confirmation` joins what waits. Tasks marked `ask` are listed with their cost. `confirmed: true` only after the merchant's yes in this chat, never because a task is `auto`; never a second scheduled task, never anything live.\n- The routine is the assistant's working day, not a list of proposals. The merchant reads, quickly: what matters today and why, what Recomaze already did, what complete work is ready for their decision, what is blocked and what is missing, and what moved since last time. Each fact in the summary must change a priority or a decision; missing access is never zero, old data never reads as new, and on the first run it is the starting point. Choose the interventions from `today_batches` and the sections by evidence, urgency, expected effect and what moves something already started: the order of the list is not a priority and sales never comes first by rule. Several independent interventions can be justified on one day, with no quota and nothing added to fill the report. Each carries its objective, why now, and every target's complete text and destination, read from `read_first` (`get_lead_detail` for a follow-up), never cut to fit and never offered for later. Each goes on one `show_card`: a row for each target with its text as the hint, with Approve, Review and Later.\n- One yes per intervention: before its card, call `approve_work` with the exact calls Approve would make and the ceiling; on the click, call it again with `confirmed` true, run those calls with the same arguments and without asking again, read each result back and report what finished and what failed. When the merchant unchecked rows, the click is their yes for the rows left: call `approve_work` for those calls alone, then again with `confirmed` true. `approve_work` takes only `apply_product_content`, `create_text_entry`, `generate_article`, `post_lead`, `post_plan`, `post_task_complete`, `write_social_post`; any other call on a card, such as an article edit, never goes in it and keeps its own yes. A tool whose arguments take `confirmed` is called before the card without it, which answers with its cost and `needs_confirmation` and runs nothing, and again with `confirmed` true on the click; a tool whose arguments have no `confirmed` is called once, only after the click, with exactly the arguments the card showed. Any other call, or a change to one, still asks; publishing and messages to customers are never part of it. End each run with `save_brief`, each item with its id as `key` and where it lands as `target`, and read `last_brief` first: continue what was approved or unfinished, respect what was put off, and never offer the same thing again without something new.\n- What waits for a yes goes on a card with `show_card`: the buttons Approve, Review and Later on one, and on several a row for each and one button for the checked ones. The verified cost sits on the button that spends it, such as \"Generate draft · [the price] credits\", never in a paragraph, and is not repeated after a yes. Never a numbered list, and never write a button as text.\n- The summary is 150 words at most besides the prepared work, measured, not estimated: before sending, pass the brief without that work to `check_brief`; when it answers `within_limits` false, leave out the least useful lines, never the prepared work, and check again.\n- Where the brief appears is the app's: a task that starts a new chat for each run posts there, so the first run's note never promises this chat unless the task says so.\n- The writing-plan question of a first run is asked only while the Writing plan line says it was not asked yet; the answer goes on that line.\n- A section that could not be read is named, not counted as zero.\n"
}