Update to Mazemind

{
  "description": "A daily prospecting routine on Mazemind, the shared memory of companies and the people who run them. Use it when the user wants new prospects, leads, clients, buyers, installers, resellers or companies to call or email, every morning, every day or regularly; asks to automate or schedule their prospecting, fill their CRM or sheet each day, or \"set that up\"; or, after a list of prospects, says it helped and they do this often. Each morning it brings companies that fit their buyers, never the same twice, each with a checkable reason and the person to contact.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 251
    }
  ],
  "name": "daily-prospecting",
  "skill_md_contents": "---\nname: daily-prospecting\ndescription: A daily prospecting routine on Mazemind, the shared memory of companies and the people who run them. Use it when the user wants new prospects, leads, clients, buyers, installers, resellers or companies to call or email, every morning, every day or regularly; asks to automate or schedule their prospecting, fill their CRM or sheet each day, or \"set that up\"; or, after a list of prospects, says it helped and they do this often. Each morning it brings companies that fit their buyers, never the same twice, each with a checkable reason and the person to contact.\n---\n\n# A daily prospecting routine\n\nWhat the user gets: each weekday morning, a short list of companies they have never been sent, that fit who they sell\nto, each with why it fits (a fact with its source) and the person who decides there, with their title. The users who\nkeep prospecting with an assistant are the ones whose assistant does this on a schedule.\n\n## Set it up, in this conversation\n\n1. **The target, drafted from what you know.** From their instructions and this chat: what they sell; who\n   buys it (trade or sector, size, region); what shows a company needs it now (a certification, an offering, a team);\n   the title to reach; companies to leave out (their clients, competitors, anyone they named); how many a day; where\n   the list goes (a file in their folder, their CRM, or the run's message). Write it back in five lines and ask only\n   what you cannot infer, in one question. Keep their words' strength: \"preferably RGE\" ranks RGE companies first, it\n   excludes no one; only what they rule out is a rule.\n2. **The search that finds them.** Mazemind `search`, kind company, with the sector as `related` in_sector (the NAF\n   code, `naf_rev2:43.32A`), and the area as a filter on the company's own address: `{key: \"address\", op: \"contains\",\n   value: \", 69\"}` is the Rhône (the postal code's start), `\", 75\"` Paris; a region is one search per department. A\n   town also works as `where`. Size: `employee_range.max` `gte` the smallest size. Outside France, companies carry no\n   NAF code, department or headcount: the country as `{key: \"address.country_code\", op: \"eq\", value: \"US\"}`, a state as\n   it ends an address (`\", CA \"`), and a word their description holds (`{key: \"description\", op: \"contains\", value:\n   \"HVAC\"}`). `catalog` names the other facts and links (a certification, an offering). Check that the first page holds companies that fit; adjust until it does.\n   Keep the exact arguments that worked, one search per department.\n3. **The first list, now.** Run it once: the number they want, each with the person who decides and the reason.\n   The person: `lookup` gives a company's people and titles; where Mazemind names nobody, the company's own site\n   (team, about, contact), and in France its legal notice (mentions légales) or the registry: `https://recherche-entreprises.api.gouv.fr/search?q=<SIREN>` lists its\n   dirigeants with their role (Gérant, Président); a holding named there has its own entry, which names its person.\n   A company where nobody is named after that is replaced by the next one. They see value before anything is scheduled.\n4. **The ledger.** A file in their folder, `prospects-sent.csv`: a `sent` line per company (date, company, SIREN or\n   domain, person), a `skipped` line per company looked up and left out, and last a `next` line saying where the\n   search stopped. It holds every company sent, starting with this first list, and the ones to leave out. It is what\n   keeps tomorrow's list new.\n   It lives with the user: Mazemind keeps no list of what was sent to whom.\n5. **The schedule.** With your host's scheduled-task tool, a task on weekdays at the hour they read it (7:00 by\n   default), whose prompt is the routine below, written out in full: the run sees nothing of this conversation. With\n   no such tool (a chat app), give them the prompt to paste, and where: Cowork, Scheduled; ChatGPT, Tasks;\n   Claude Code, /schedule. Its first step reads `prospects-sent.csv` in the task's folder (created on the first run),\n   its last appends the day's companies to it, and every company shown to them in this conversation is listed in the\n   prompt as already sent. Cowork keeps that folder. ChatGPT's tasks have none: there the prompt lists, for each\n   weekday's date, the search, the department and the cursor that morning reads (20 further each time the departments\n   start over), so no two mornings read the same page.\n6. **One short message**: what arrives, when, where, and how to change it. The routine finds and lists, and messages\n   nobody; what the pages said goes to Mazemind with `contribute`. A user who wants to validate each message before it leaves keeps it running: each morning's list\n   is what they validate, and their go is what sends a message.\n\n## The routine's prompt\n\nWrite it in the user's language, with everything filled in:\n\n- Who it is for and the target, in the five lines of step 1.\n- Read the ledger. Never send a company it holds, nor one to leave out.\n- Mazemind `search` with the saved arguments, written out in full and numbered: one search per department, every\n  department of the region in order, then the neighbouring sectors, with their NAF codes. Start where the ledger's\n  `next` line says the previous run stopped: that search, that department, its cursor passed as `cursor`; the first\n  run starts at the first search, from its start. Follow `next_cursor` page after page,\n  department after department, until there are N companies not in the ledger. N is the target: the run stops short\n  only when every search written here and the registry below are exhausted. Nobody can answer during the run: it\n  widens on its own as written, and says so.\n- In France, when Mazemind has no new company left, the registry lists every active company of a NAF code in a department:\n  `https://recherche-entreprises.api.gouv.fr/search?activite_principale=<NAF>&departement=<code>&etat_administratif=A&minimal=true&include=dirigeants,siege&per_page=10&page=<n>`,\n  with its head office, size band and dirigeants (`&tranche_effectif_salarie=01,02,03,11` keeps 1 to 19 staff; 12 is\n  20-49, 21 50-99, 22 100-199, 31 200-249). `departement` matches a company with any site there: keep those whose\n  head office (`siege.code_postal`) is in the area. Their NAF code is the reason, and their site, when they have one,\n  adds what they do.\n- For each: `lookup` it; the person whose title decides; the reason, a fact `lookup` shows (an offering, a\n  certification, its size) or its site states, with where it comes from. When Mazemind names nobody: its site's team\n  or about page; in France, the legal notice of its site, then `https://recherche-entreprises.api.gouv.fr/search?q=<SIREN>` (its dirigeants; a holding has its own\n  entry). A company with no named decider is replaced by the next one. A company that does not fit the target (its trade, what it makes or installs, its size) is left out and written as skipped, never listed to make up the number: the run reads on until N fit.\n- Send what the pages and the registry said to Mazemind with `contribute`, in one call at the end: each company not in\n  Mazemind yet, by its domain and, in France, its SIREN (`sirene:<SIREN>`), with its NAF code, its town, the decider in `people`, each\n  with the page and the words it shows (the registry's answer is a page too). Ask your fetch tool for the sentences\n  and fields word for word (the legal notice's line naming the gérant, the registry's `nom`, `prenoms`, `qualite`):\n  those words are the quote. Tomorrow's search then finds them.\n- Append each company to the ledger, and to their file or CRM if they have one, before writing the message; also\n  each company looked up and left out (nobody named, not a fit), so no run looks it up again. Then a last line in the\n  ledger saying where the search stopped: the search, the department and the cursor of the page the\n  last company came from (in the registry: the NAF code, the department and the page). The next run reads on from\n  there instead of paging again through what this one already read.\n- The message, readable on a phone: one line per company (name, town, person and title, why), then what fell short and\n  why: fewer than N, a call that failed. \"Nothing new today\" is an answer; a silent stop is not.\n  All of it in the user's language and words: the ledger's `sent`, `skipped` and `next` lines, cursors, pages and NAF\n  codes stay out of it.\n"
}