Write a different weekly update for everyone

9 min read · the layout, the prompt, and all nine rules · every claim sourced below

The short version: you can have your computer write a separate weekly message for every person on your list, ready before you get up on Monday. It takes one spreadsheet and Claude Code, and the whole build is below: the exact sheet layout, the prompt to paste, and the rule list that is doing the real work. A marketer at Anthropic runs this for his whole team, and by the end of his first week his prompt held nine rules, each one added because somebody complained. All nine are here.

Build it yourself

Two things, and that is the whole list: a spreadsheet, and Claude Code — which is AI you point at your own files, running on your own machine. No website, no hosting, and nothing to code.

The version in Anthropic’s write-up reads from a company data warehouse, and that is the least transferable part of it. What transfers is the shape: one source of truth, a message per person, and a rule list that gets longer every time the output is wrong.

Step 1 — the spreadsheet

One file, two tabs. Nothing else. Column names are yours to change; what matters is that the first row is the header row, because the prompt reads it.

The interested_in and who_it_suits columns are the whole matching mechanism. Keep the words identical between the two tabs — that is what stops a retail account being invited to a finance dinner, which is rule 4 below and a mistake that has actually happened.

A row with no link is a row nobody can act on. Leave the cell empty and the prompt below drops it, which is rule 2.

Step 2 — write one message yourself

This is the step everyone skips and it is the highest-leverage one on the list. Before connecting anything, write a single message by hand, with invented content, exactly as you would want it to land. Save it as example.md next to the sheet.

Ward did this at Anthropic before touching any data, and wrote a second one for managers because they want a view of the whole team rather than a list of individuals. A made-up example settles format, length and tone in one go, and it costs ten minutes.

Step 3 — the prompt

Put the sheet and the example in a folder, open Claude Code in that folder, and paste this. Edit the bracketed parts.

Read the output. Every time one is wrong, add a numbered rule under HARD RULES saying what should have happened — not a vague preference, a rule. That is the whole method, and it is the only part of this that actually determines whether the finished thing is any good.

Two details worth keeping from Ward’s own advice. Version the prompt file, with a numbered version and a one-line note of what’s changed, so you can see which rule broke something. And write the instructions in plain language — the rules above are sentences, not syntax.

When the drafts are consistently right, and only then, connect the sending. Not before. The last line of that prompt is do not send anything for a reason: an automation that messages your colleagues is a bad place to discover your column map was wrong.

What the job actually is

Start with the person, because the technology makes no sense without him. Adam Ward works on Anthropic’s marketing team, in a role called field marketing. His job is to make sure the people selling the product know what marketing is doing: which events are running, which webinars are open, what new material exists that a customer might want to read.

The way he used to do that was completely manual. On Sunday evening he would sit down and build a slide deck consolidating the week’s business updates. On Monday morning he presented it at the team stand-up, then dropped it into Slack for anyone who missed it.

It worked until it did not. As the team grew and he picked up more sales teams, the first thing to break was not the deck — it was the personalisation. In his words, he no longer had time to pick out the opportunities that were right for each team. One deck for everybody means most of it is irrelevant to any individual reading it.

What arrives instead, now

Nothing gets presented. Every Monday, each account executive gets a direct message in Slack containing five things:

1. Three priority actions for the week, as a “top three things” list at the top.
2. Field events that relate to their own accounts.
3. Contacts of theirs who have already registered for upcoming webinars.
4. Marketing content worth sharing — blog posts, ebooks, customer stories.
5. Follow-up suggestions.

The important word is their. Each message is composed from the recipient’s own account list, so no two messages are alike. That is the thing the Sunday deck could not do at scale, and it is the entire reason the project exists.

How it was built

The first version came out of an internal marketing hackathon. The sequence is worth following exactly, because the order is what kept it small enough to finish.

The nine rules, in full

Here they are. Anthropic’s write-up says all nine trace back to a piece of feedback from a seller or a manager, which means you can read the list in two directions: forwards as instructions, and backwards as a list of everything the first version got wrong.

The right-hand note on each is what that rule prevents. Where the write-up names the specific complaint that caused it, I have said so; where it does not, I have not guessed.

1. 1. Never invent a URL. A link renders only if the address comes character for character from the source sheet.

   Copy

   The tell. You do not write this rule unless links were appearing that did not exist. This is the single most transferable line in the whole prompt: any AI assembling a message from a data source needs it.

2. 2. Drop events with no link entirely.

   Copy

   Their reasoning, quoted: events for which our sellers can't register anyone are just noise. The earlier version padded the message with things nobody could act on.

3. 3. Check contact titles against the event's intended audience, and drop mismatches without comment.

   Copy

   Added after a seller flagged an engineering VP being recommended a workshop aimed at knowledge workers. Note the second half: dropped silently, not flagged, because a message full of apologies for its own filtering is worse than a shorter message.

4. 4. Apply an industry gate before recommending an event.

   Copy

   The published example is keeping retail accounts off finance dinner invitations. Relevance is not just about the person, it is about the account they sit on.

5. 5. Brand-new sellers with no accounts yet get a short welcome note, not a blank message.

   Copy

   Somebody's first week at the company ended with an empty automated message. This is the rule that costs nothing and buys the most goodwill.

6. 6. Open every run by reading the sheet's header row and verifying the column map before composing anything.

   Copy

   The field events sheet had its columns rearranged three times in six weeks. Without this, the first Monday after someone inserts a column, every message quietly contains the wrong field. Steal this one even if you steal nothing else.

7. 7. Managers get a separate roll-up format, not the individual message.

   Copy

   Written in from the start, because managers want a holistic view of their team rather than a list of individual accounts. Two audiences, two templates.

8. 8. Open with a fixed 'top three things for the week' list of three action items.

   Copy

   A consistent opening is what makes a recurring automated message readable at a glance. The reader learns where to look, and the format never argues with them.

9. 9. Pull the rep's territory from the CRM, plus any relevant account updates.

   Copy

   The personalisation input. Territory decides which events are even geographically plausible, and it is the field that made extending to business development reps a one-field change.

Did it work?

Two claims are published, and they are worth separating.

The soft one is the scale. It went from one pilot team to every team he supports; field marketing now runs it for all of sales; business development reps were added by changing a single field in the prompt; customer success and alliances followed. Every send is archived in full, so managers can read what their team was told.

The hard one is a single number: registrations for an executive dinner doubled in a week, which he attributes to the right reps having the right event in front of them on Monday morning. Take that for exactly what it is — one event, self-reported, no control group. It is the only outcome figure given.

The claim I find most convincing is neither of those. He went on holiday, and the Monday send went off on its own, without a hitch. An automation nobody has to be awake for is a different category of thing from a script you run.

The honest bit

Three things this story is not.

It is not a secret anybody discovered. This is Anthropic’s own published case study about their own employee using their own product, and it is written to make Claude Code look good. That does not make any of it false — the rules are specific enough to be useful precisely because they are unflattering — but it is marketing, and you should read the nine rules rather than the adjectives.

It is not free. Claude Code is a paid product, and a job that runs weekly across a whole sales organisation is not a trivial amount of usage. The write-up gives no figure for what it costs to run, so neither will I.

And it is not a weekend project at that scale. The version in the article reads from a company data warehouse, which someone had to set up and secure. The one-spreadsheet version above genuinely is an afternoon. The gap between them is not prompt engineering, it is plumbing.

If you want the underlying idea rather than this one implementation: a rules file that persists between runs is the same mechanism as a project config file, and correcting the output until it holds is the manual version of a self-correcting loop. This story is what those two look like pointed at a job that repeats every week.

The one line to take away

Before composing anything, read the sheet's header row and verify the column map. Never assume a column position.

Columns move. People insert things. That one line is the difference between an automation that fails loudly and one that quietly sends two hundred people the wrong field for six weeks.