business-os

Build a business owner's AI workspace from nothing, in one warm conversation: research their website and Instagram first, interview them one question at a time about the business, its two audiences, brand voice, offers, rules and how they want to be worked with, then create their Google Drive workspace and every document in it, set up their publishing accounts, and draft three real posts so they see it working the same day. Use when an owner is new or says "set me up", "get started", "onboard me", "connect my accounts", "you don't know my brand yet", or asks what this can do — and when any other skill finds no workspace in Drive at all. Use it again only to re-run a whole section of the interview from scratch. For changing something that already exists, use update-the-brain instead.

Write one Instagram + Facebook post in the owner's voice from a Content Calendar row or a one-line idea: hook in the first 125 characters, one clear call to action, local hashtags, first-comment text, alt text, and — when the row calls for a graphic — hand off to make-graphic. Runs the guardrails before saving and stages the result as a post packet in 2 Drafts for approval. Use whenever the owner says "draft this", "write a post about…", "make the posts for this week", "turn this into a caption", or plan-week hands over rows. Never publishes.

The monthly maintenance pass and safety net that keeps the owner's AI workspace healthy for years instead of months. Audits the invariants the change-time guard is supposed to hold — no two documents doing the same job, the map matching what is really in Drive, nothing edited by hand behind our back — then archives published content that has aged out, compacts weekly learning notes into quarterly summaries, flags brain documents that have gone stale, surfaces accounts needing reconnection and posts nobody retried, and reports loose files that need a home. Use when the owner says "tidy up", "housekeeping", "is everything okay", "spring clean", "my folder is getting messy", or when a monthly scheduled task runs. Reports and proposes; never reorganises her folder without a yes.

The weekly learning loop that makes next week better than this week: pull real Instagram and Facebook performance from the publisher, find what worked and why, detect failed posts and accounts that need reconnecting, check the best posting times, write a plain-English report to 3 — Social/Results, and update Learned by us / What works (current) so plan-week and draft-post read it next time. Use whenever the owner says "how did we do", "what's working", "weekly report", "any problems", or a weekly scheduled task runs. Read-only against the publisher except for retrying failed posts with permission.

Create an on-brand social graphic in Canva for a post — carousel slides, tips cards, quote cards, availability announcements — using the owner's colors, fonts and voice from their Drive brain, sized for Instagram and resized for Facebook, and hand back a Canva design id plus an edit link for the post packet. Use whenever a post's media source is "canva", the owner says "make a graphic", "design this", "turn this into a carousel", "I need an image for this post", or draft-post asks for one. Works on Canva Pro; does not require Enterprise. Does not export the final file — publish does that at approval time.

Plan the coming week of Instagram and Facebook content for a business owner: read their brain (voice, offers, audiences, what has worked), pick a balanced mix across topics and both audiences, choose formats and posting times from real performance data, and write it into the Content Calendar so drafts can be made. Use whenever the owner says "plan my week", "what should we post", "content for next week", "I've gone quiet", "fill the calendar", or a weekly scheduled task runs. Also use when someone asks for a month of content — plan it week by week. Never publishes anything; it only plans.

The only skill that talks to the publishing vendor. Turns an approved post packet into a scheduled Instagram and Facebook post: exports the Canva design if there is one, resolves media, validates it, checks the Instagram daily quota, creates or promotes the Zernio post with the exact publish time, records the vendor id, and moves the packet through 3 Approved → 4 Published. Also hands the owner an upload link for her own photos and Reels, retries failed posts, and reports status. Use whenever the owner says the approval word ("approve <post>"), "schedule this", "post it", "what's scheduled", "did it go out", "retry", "reschedule", or "cancel that post". Never publishes without the owner's explicit approval in the current conversation.

Audits whether a draft reads as though a person wrote it, and rewrites the shape when it doesn't. Runs the structural pass that word-level editing cannot reach — a lesson stated too often, an arc that resolves too neatly, emotion performed through the body instead of named, vague allusion where a real thing should be named, and a skeleton identical to the last three posts — then checks the result still sounds like the owner rather than merely clean. Use when she says "this sounds like AI", "this doesn't sound like me", "make this sound human", "humanize this", "it reads robotic", "too polished", when a draft is for something longer than a caption, or before anything goes out that a reader will judge on trust.

Learns a job the owner already does, by interviewing her about it rather than guessing how she works, and writes it down as a playbook in her own workspace so it runs the same way every time. Use when she says "can you do X for me", "I want you to handle my invoicing", "let me show you how I do this", "here's my process", "do it the way I do it", when she describes a recurring task this system does not already have a job for, or when a shipped job keeps being corrected in the same direction because her real process is different from the default. Extracts the trigger, the inputs and where they live, the steps in order, the decisions and their reasoning, what done means, the edge cases, and what a bad version looks like — then writes the checks before the job ever runs.

The only safe way to change what this system knows or how it behaves, and the required first response when the owner pushes back. Use when she corrects something ("that's wrong", "no, we don't do that", "I never said that"), asks why something happened ("why did you post that", "where did that come from"), repeats herself ("I already told you", "we talked about this"), disagrees or sounds frustrated, changes a fact ("we charge more now", "we've added overnight care", "change my brand voice"), asks to add a room, connector or new kind of work, or when you notice two things in her workspace that disagree. Finds the document that caused the behaviour, fixes that document, fixes everything built from it, and records it — instead of apologising and leaving the old version in place.

Brings an owner's workspace up to the current shape when the system has changed since it was built, and repairs one that is out of shape. Runs two ways: quietly filling in anything missing (a folder, a document, a map row) without asking, and — for anything that renames, moves, merges or retires — showing her exactly what will change before it changes, then doing it on one yes. Use at the start of any session where the workspace version is behind, when a document or folder the map promises is not there, when the map disagrees with what is actually in Drive, when a previous upgrade is recorded as FAILED, or when the owner says "is my setup current", "something's missing", "this looks broken", or "update my workspace". Never deletes, always resumable, safe to run twice.