Cookbook
The work your team does every week, done without them
Every recipe below used to be somebody's Tuesday morning. Now it runs on a
schedule, or the moment the mail lands, and nobody opens it. You do not configure
any of them. You describe the routine in your own words, an agent assembles the
workflow while you watch, and it hands you the short list of things only you can
connect: your mailbox, your Slack, your repository.
Each recipe ends with the words to paste into a chat with your agent, the same box
you would type a message into. It builds the workflow while you watch, and
anything that sends in your name can wait for your approval first. Change the
names to yours.
Every recipe puts something in front of a person: a message in your chat, a row in
a document, a page at a link, a draft waiting for you to send. Anything that sends
in your name can be set to wait for your approval first, and every run keeps a
record you can open afterwards, including the ones that got it wrong.
The hour a day that goes to the inbox
Work arrives as mail whether or not anyone wanted it to. These three read it, sort it and draft the replies before you sit down.
Sort the inbox before you open it
You open a sorted inbox instead of a pile. Everything that came in overnight has already been read and labelled: what needs you, what is a supplier invoice, what is a recruiter, what is noise you never have to see.
How it runs When mail arrives at the address you name → decide what kind each one is → post the customer ones to your chat, file the rest
Ask for it
When mail arrives at ops@, decide whether it is a customer question, a supplier invoice, a recruiter, or noise. Post anything that is a customer question into the Support chat with a one-line summary, and ignore the noise.
Invoices into a spreadsheet
Nobody retypes an invoice again. Each PDF is read as it arrives and the supplier, date, net, VAT and total come out of it into a new row in your Invoices document, ready for whoever does the books.
How it runs When mail arrives with a PDF attached → read the attachment → pull out the fields you named → add the row to the Invoices document
Ask for it
Whenever an email arrives with a PDF invoice, pull out the supplier, invoice number, date, net, VAT and total, and add them as a row to the Invoices document. If a field is missing, say so in the row rather than guessing.
Drafts for the ones that need a reply
The dull half of answering mail, digging out the thread and remembering what was agreed in March, is done before you sit down. The draft is waiting in your house style, and nothing is sent until you send it.
How it runs When a customer emails → check what the workspace already knows about them → draft the reply → post it to you unsent
Ask for it
For each customer email, search the Support notes for anything we have already told this customer, then draft a reply in our house style and post it to me for approval. Never send it yourself.
The conversations that repeat every week
The same handbook question for the fourth time this week. The same meeting, the same three answers. None of it needs you in the room, and all of it happens in a chat window people already have open.
A Telegram bot that answers from your documents
Your handbook, your price list, your policies, answerable in a chat by anyone you hand the bot to, without them reading any of it. Correct an answer once and that is the answer everybody gets from then on.
How it runs When someone messages the bot → find the answer in what you have taught it → reply in the chat, quoting the passage it came from
Ask for it
Make a Telegram bot. When someone messages it, search the Handbook memory for the answer and reply in two sentences, quoting the passage it came from. If the handbook does not cover it, say so instead of improvising.
The status meeting that became a message
Everyone answers in their own time, in the chat they already have open. You read one summary instead of putting six people in a room for fifteen minutes, and whoever is blocked is at the top of it.
How it runs Every weekday at the time you set → ask each person in the channel → wait for the replies → post one summary
Ask for it
Every weekday at 9:30, ask each member of the Engineering channel in Slack what they finished yesterday and what is blocking them. Collect the replies until 10:00, then post a single summary and list anything that is blocked.
A Discord bot for the questions asked weekly
The handful of questions your team answers every week, answered the same way every time, at three in the morning, in the voice you chose. When the answer changes, you change it once.
How it runs When a question is asked in the channel → answer from what you have taught it, in the voice you set
Ask for it
Answer questions in the #help channel from the FAQ memory scope. Keep replies short and friendly, link to the docs page when there is one, and tag a human if the question is about billing.
Watching things so you do not have to
Pages change, feeds fill up, a competitor launches something. None of it needs your attention until it does.
A digest of what changed overnight
Release notes, changelogs, a competitor's blog - read every morning, condensed to what actually matters to your team, waiting before standup.
How it runs Every morning before the day starts → read the pages you list → write up what changed → post it to your chat
Ask for it
Every morning at 07:00, fetch the changelog pages I list, summarise what changed in a way our support team will care about, and post it to the Release digest chat.
Tell me only when it matters
The thing you keep refreshing by hand, checked on whatever interval you name, as often as every minute. You hear about it when it crosses the line you set, and never while it is fine.
How it runs On the interval you name, as often as every minute → check the page → message you only when it crosses the line you set
Ask for it
Check this status page every hour. Tell me on Telegram when the number waiting goes over fifty, and do not tell me again until it drops back under.
A feed, read for you
Twenty industry feeds become five things worth knowing, once a day, in your own words.
How it runs When the feeds you follow publish → drop the kinds you said you do not care about → one line each on the rest, once a day
Ask for it
Follow these RSS feeds. Once a day, drop anything that is a press release or a funding round, summarise the rest in one line each, and post the five most relevant to us.
The reading nobody has the afternoon for
The contract, the hour of transcript and the four regional spreadsheets all need somebody to read them properly, this week, again. What you looked for last time is what gets looked for this time, without you having to say so again.
Check a contract against your own checklist
The clauses you always look for, looked for every time, with the page they were found on or a clear note that they are missing. Anything you add to the checklist is looked for in every contract after it, including the clause you negotiated away in the spring.
How it runs When a contract is added → check it against your checklist → report what is present, missing or unusual
Ask for it
When a contract is uploaded to the Legal scope, check it against our review checklist and tell me which clauses are present, which are missing, and quote the wording for anything unusual.
A report out of a folder of spreadsheets
You stop chasing four regions for their spreadsheets. When the last one lands, the report writes itself in the format you use every month, with the total and the three biggest movements at the top.
How it runs Watch the folder → wait until every regional sheet is in → combine them → write the report the way you write it
Ask for it
Watch the Monthly numbers folder. When all four regional spreadsheets are in, combine them, write the summary in the usual format, and post it with the total and the three largest movements.
Meeting notes into decisions
An hour of transcript becomes the four things that were decided and who owes what.
How it runs When a recording is added → transcribe it → pull out the decisions and who owes what → post them to your chat
Ask for it
When a recording lands in the Meetings scope, transcribe it, then post the decisions taken, the actions with owners, and anything explicitly deferred. Leave out the discussion.
Code, in a container of its own
The review nobody has time to do before merging, the dependency check nobody remembers, the bug nobody has reproduced yet. These run in a container of their own, separate from your computer and from everybody else's, with only the access you give it and no network at all if you say so.
A first look at every pull request
Connect the repository once. After that every pull request is cloned and the suite actually run, not guessed at, so the obvious problems are caught before a person spends attention on them.
How it runs When a pull request opens → clone the branch with the access you connected once → actually run the test suite → post the review before a person opens it
Ask for it
On every pull request, clone the branch, run the test suite, and post a review: what failed, anything that looks risky, and whether it is safe to merge.
The dependency check nobody remembers to do
Every Monday the project gets checked as it actually builds this week, for dependencies that are out of date or insecure. Anything critical reaches you that day. Everything else waits for the weekly digest.
How it runs Every Monday → clone the project → run the check against what it actually builds → message you that day if it is critical, otherwise add it to the digest
Ask for it
Every Monday, check the project for outdated and vulnerable dependencies. If anything is critical, message me that day. Otherwise add it to the weekly digest.
Reproduce the bug before anyone reads the ticket
The bug is waiting on your desk with the steps already tried and the output attached. Triage starts from evidence instead of from somebody asking whether it happens on your machine too.
How it runs When a bug is filed → try the steps in a container of its own → capture the output and any error → post the attempt back onto the ticket
Ask for it
When a bug is filed in the tracker you connected, try the reproduction steps in a container of its own, capture the output and any error, and post the result back onto the ticket.
Stop being the person who exports the spreadsheet
A number nobody can find is a number nobody uses. These end in a page with its own address that you can send to anyone, including the people who will never have a login here. The figures come from wherever they already live: a nightly export dropped into a watched folder, or your own systems posting them in as they go.
A board for the numbers you keep being asked for
Sales this week, open tickets, what is overdue, on one page with its own address, refreshed while you sleep. You send the link once and the four people who keep asking you stop asking you.
How it runs Ask an agent to build the page → it serves it at an address of its own → a schedule refreshes the figures every morning
Ask for it
Build me a small dashboard showing this week's orders, open support tickets and anything overdue, reading from the Ops memory our order system posts into each night. Put it at an address I can send round. Then set up a workflow that refreshes the figures every morning at 06:00.
A status page your customers can see
The things you already check by hand, checked every few minutes and shown publicly, so the next outage does not also mean answering the same question forty times.
How it runs Every few minutes → call each endpoint → write the result onto a public page → post to your chat only when something changes state
Ask for it
Every five minutes, check these endpoints and record whether each answered and how quickly. Put up a plain status page showing each one as up or down with the time of the last check, and post to the Ops chat only when something changes state.
The Monday report, as a page instead of an email
The same weekly write-up, but it lives at a URL, keeps its history, and nobody has to dig through their inbox for the version from three weeks ago.
How it runs Every Monday morning → pull last week's figures → write the report onto its own page → post the link to your chat
Ask for it
Every Monday at 08:00, pull last week's figures, write the summary the way I usually write it, put it on a page titled with the week's dates, and post the link to the Leadership chat.
A wall board for the room
The screen nobody has to touch: today's bookings, the deploy queue, who is on call. It reloads itself and it is never out of date.
How it runs Ask an agent to build the board → it serves it at an address you open on the screen → a schedule keeps the numbers current
Ask for it
Build a full-screen board for the office TV showing today's bookings, what is deploying and who is on call. Big type, nothing to click, refreshing itself every minute. Give me the address to open on the screen.
The things you meant to do regularly
Everybody has three of these. They get done when somebody remembers, which is not every week.
The morning briefing
Your calendar, your inbox, what broke overnight and what is due - one message, before you start.
How it runs Every weekday before you start → gather your calendar, your inbox and the overnight alerts → post one message
Ask for it
Every weekday at 08:00, post a briefing: what is on my calendar, anything in the inbox that looks urgent, and anything from the overnight alerts that is still unresolved.
Chase the invoices politely
The follow-up you keep not sending, sent on time, in the right tone, and stopped as soon as they pay.
How it runs On the first of the month → find what is overdue → check whether it has since been paid → email only the ones still outstanding
Ask for it
On the first of each month, find invoices more than thirty days overdue, check whether payment has since arrived, and email a polite reminder only to the ones still outstanding.
Prove the backup restores
Not that the backup ran, but that it can be restored. Once a week the real file is restored into a container of its own and the numbers checked, and you hear about it only if something is wrong.
How it runs Once a week → restore the backup into a container of its own → check the row counts look right → tell you only if they do not
Ask for it
Once a week, restore the latest backup into a container of its own, check the row counts look sane, and tell me only if something is wrong.
Yours is the one that eats a morning every week
That is the usual case, and it is why nothing here is a template. Describe the
routine you actually do, in the words you would use to explain it to somebody
new, and an agent assembles the workflow and tells you which accounts it still
needs before the first run. If it cannot do a part of it, it says so rather than
pretending.
One price, your own subdomain, model access included.