Personalized weekly digests
A recurring update loses its value when there is no time to personalize it. The pattern in this skill turns a hand-built briefing into a system that composes one message per recipient, and it treats every correction from a recipient as a rule that gets written down rather than remembered.
The approach comes from a field marketer at Anthropic who replaced a Sunday-evening slide deck with an automated Monday-morning digest for every sales rep he supported. The first version took about an hour to build; the value came from the weeks of pilot feedback that followed.
Instructions
1. Brief the business problem, not the implementation
You do not need coding skills to build this. You need to explain the problem clearly.
- Open by describing your challenge and positioning yourself as the product manager of the output, not the technician building it.
- Give business context the way you would brief a new colleague. Recording a voice explanation of the problem is an effective way to get the full context across.
- Use templates/kickoff-prompt.md as the opening brief.
2. Show the output format with a template example
Abstract instructions about tone and structure underperform a concrete example of what a good message looks like. Provide a filled-in sample.
- Structure the message around the "top three things for the week," prioritizing items the recipient can act on.
- Use templates/digest-message-template.md as the shape, and fill it in once by hand so Claude has a target to match.
3. Connect the real data sources
Personalization comes from joining the recipient to the data, not from generic phrasing.
- Connect Claude to your marketing/CRM warehouse over MCP. In the source case that was BigQuery, which itself pulls from HubSpot, Clay, and Salesforce.
- Pull the recipient's territory from the CRM, relevant account updates from Slack, and match those against the current marketing initiatives.
- Expand the source set over time — blog articles, eBooks, customer stories, webinars, and partner ecosystem events all became part of the briefing.
- See references/data-sources.md for the wiring and for the schema-drift instructions below.
4. Make the prompt survive schema drift
Source spreadsheets change without warning. The field events sheet in the source case rearranged its columns three times in six weeks.
- Instruct Claude to read the header row and verify the column mapping before processing rows.
- Refer to columns semantically — "look at the column with the event URL" — instead of hard-coding column positions or letters.
5. Pilot with a committed group and route early runs to yourself
- Pick one team small enough to give real feedback and willing to invest in it; the source pilot was a single sales team of ten reps.
- Send early runs to yourself first and judge the output against the quality baseline you already hold from doing the task by hand.
- Only widen distribution once the obvious data-quality and relevance failures are gone.
6. Convert every correction into an explicit rule
This is the core operating loop. Each piece of feedback becomes a documented rule, and you track which rule came from which feedback. After the first week the source prompt held nine explicit content rules.
Rules that emerged in the source case:
- Never invent a URL. Claude had produced plausible-looking URLs for events that were missing registration links. Render links only from exact source data.
- Verify contact titles against the event audience. Engineering VPs were being invited to knowledge-worker workshops.
- Gate by industry. Retail accounts should not receive finance-focused events.
- Handle the empty case. New sellers without assigned accounts were getting blank messages; they need a custom welcome note instead.
The full catalog is in references/content-rules.md, and worked conversions from raw feedback to rule text are in examples/feedback-to-rule.md.
7. Version the prompt as it grows
- Save each update as a numbered version with a one-line change note.
- Start in a document, and move to a collaborative platform such as GitHub as more of the team needs access.
8. Scale by duplicating, then changing the mapping field
Each new audience is a copy of the working prompt with the one field that describes how that audience maps to accounts in the CRM changed. In the source case, a BDR version launched two days after the request; customer success, alliances, and cross-functional partners followed.
9. Run it as an operational system
- Archive every send for audit and accountability.
- Give managers a consolidated roll-up of their team's recommendations.
- Move from approving each send to reviewing output, so the system keeps running when you are away.
Examples
Framing the opening brief
I'm a field marketer supporting three sales segments. Every Monday each rep needs to know the three things that matter for their accounts this week. I'm the product manager for this output, not the engineer — here's the problem, and here's what a good message looks like.
Attaching a filled-in sample message alongside this framing does more than any amount of description of the desired tone.
A rule written from a single piece of feedback
Feedback: "You sent me a registration link that 404s."
Rule added to the prompt:
Never invent a URL. Only render a link when the exact URL is present in the source data. If an event has no registration link, name the event and omit the link rather than constructing one.
A column reference that survives a reshuffled sheet
Instead of "column D holds the event URL," write:
Read the header row first and map each needed field by its header name. Look at the column with the event URL. Confirm the mapping before reading any data rows; if a required header is missing, stop and report it rather than guessing.
Standing up a second audience
The account-executive prompt already works. For BDRs, duplicate it and change only the field describing how BDRs map to accounts in the CRM — BDRs map differently from AEs. Everything else, including the nine content rules, carries over unchanged.
Source
- https://claude.com/blog/how-an-anthropic-field-marketer-uses-claude-code-to-send-weekly-personalized-updates-to-every-sales-rep (published 2026-08-24, by Adam Ward)