The stack
- Tooling
- a Claude plan with Routines + the MCP connectors the task needs
- Per run
- a few minutes of agent time, cents of usage
- Setup
- about an hour to write and tune one Routine's brief
- After setup
- runs on its schedule, unattended
The problem
Every GTM team has the same graveyard of good intentions: the weekly pipeline-risk review, the competitor check, the CRM-hygiene sweep, the 'who has not been touched in 30 days' list. Each one matters. Each one is somebody's job in theory. And each one slips the first genuinely busy week, because a chore that depends on a human remembering is a chore that does not happen.
The tell is that these tasks are perfectly specifiable. You can write down exactly what the Monday report should contain and where it goes. There is no judgment a person uniquely adds to running it; the judgment is in reading the output, not producing it. That is the precise shape of work a scheduled agent does better than a person, because the agent never has a busy week.
A Claude Routine is that agent: a scheduled cloud agent that runs on a cadence you set, pulls its inputs from the web or your tools over MCP, does the work, and drops the output in Slack or writes it to the CRM. You stop being the cron job. The recurring work shows up on time, in the same shape every time, and your attention goes to acting on it instead of assembling it.
How it works
- 01 Set the cadence Claude Routineweekly, nightly, monthly
- 02 Gather inputs Web search / MCPthe live web, CRM, docs
- 03 Apply the rules Claudematerial only, in a fixed shape
- 04 Shape the output Clauderanked, tight, scannable
- 05 Deliver it Slack / CRMwhere the reader already is
- 06 Tune, then leave it Youa couple of runs, then hands-off
- List the recurring GTM chores that keep slipping, and pick the one that hurts most when it does
- Write the task as a brief: what to gather, how to shape it, and where it goes
- Connect the inputs and outputs: web search for research, MCP connectors for your CRM and Slack
- Set the schedule (weekly Monday 7am, nightly, monthly) and let the Routine own the cadence
- Tune the output for a couple of runs, then leave it alone
- Add the next Routine, and the next, until the recurring work runs itself
The playbook
Pick the chore that hurts most when it slips
List the recurring things that are supposed to happen and reliably do not: the Monday pipeline-risk digest, the weekly competitor check, the monthly CRM-hygiene sweep, the daily new-lead brief. Pick the one whose absence you feel most, because a first win you actually notice is what earns the second and third Routine.
Sanity-check that it is genuinely a Routine and not a judgment call in disguise. If the task is 'decide our Q3 pricing', that is not a chore to schedule. If it is 'every Monday, list the deals with no next step and no activity in 14 days and post them to #revops', that is a Routine all day.
Write down what 'done' looks like in one or two lines. That description becomes the spine of the brief and the yardstick for whether the Routine is working.
TipStart with a read-and-report Routine, not a write-to-the-CRM one. A digest that is wrong is a bad Monday; an autonomous hygiene sweep that is wrong edits records. Earn the write access, same as any agent.
Write the task as a brief
Write the Routine one clear brief: what to gather, how to shape it, what to leave out, and where to send it. Treat it like briefing a reliable analyst who will do this every week without you in the room, which is exactly what it is. Be specific about the cut, because a recurring report that includes everything gets muted by week three, the same failure mode as any digest.
Encode the judgment as explicit rules. 'A deal is at risk if it has no next step and no activity in 14 days.' 'Only flag competitors with a material move this week.' The Routine applies those rules identically every run, which is the consistency a human doing it by hand never quite manages.
End the brief with exactly where the output goes and in what shape: 'post to #revops as a short list, most-at-risk first, one line each with a link to the deal.' The destination is part of the task, not an afterthought.
Connect the inputs and outputs
Give the Routine what it needs to gather and to deliver. For research, that is Claude's built-in web search. For anything in your stack, that is the MCP connectors: HubSpot or Salesforce to read the pipeline, Google Drive for docs, Slack to post. The Routine reads and writes those tools directly, so there is no orchestration platform in the middle.
Scope the access to the task. A pipeline-risk digest needs to read deals and post to one channel; it does not need write access to your contacts. Least privilege keeps a scheduled, unattended agent from being a standing risk.
If the Routine writes anything back (updates a field, creates a task), keep that behind the same care you would give any automation: start by having it propose the change for a human, and only automate the write once the proposal has been right for a while.
Set the schedule and let it run
Pick a cadence that matches the work and a fixed time the team learns to expect: Monday 7am for a weekly read, nightly for a fresh-lead brief, the first of the month for a hygiene sweep. The Routine owns the cron; you stop being it. Because it is a cloud agent, it runs whether your laptop is open or not.
For the first couple of runs, read the output like a critic and tighten the brief: cut the noise, sharpen the rules, fix the shape. This calibration is short, and it is the difference between a Routine the team reads every week and one they mute.
Once it is tuned, leave it alone. Revisit it only when the underlying work changes, a new competitor to track, a new risk rule, and otherwise let it be the boring, reliable thing it was built to be.
What you get
A starter menu of GTM Routines worth scheduling, and what each one does.
A GTM Routine menu (pick one to start):
1. MONDAY PIPELINE-RISK DIGEST (weekly, Mon 7am)
Reads open deals in HubSpot, flags any with no next step or no activity in 14 days, posts a ranked list to #revops. Manager walks into Monday knowing where to push.
2. NIGHTLY NEW-LEAD BRIEF (nightly)
For yesterday's inbound, researches each company with web search and posts a one-paragraph brief per lead to the owning rep, so first calls start prepped.
3. WEEKLY COMPETITOR DIGEST (weekly)
Web-searches your competitor list for the last 7 days, keeps only material moves, posts the 'so what for reps' to #sales. (Full build: the competitive-intel play.)
4. MONTHLY CRM HYGIENE SWEEP (monthly)
Lists records missing key fields, deals stuck in a stage too long, and likely duplicates, and posts the cleanup list to RevOps. Proposes, does not auto-edit, until trusted.
Each is one brief, one schedule, and one place the output lands.Pitfalls to avoid
Scheduling a judgment callA Routine is for specifiable, repeatable work, not decisions. If producing the output needs a human's call each time, it is not a Routine, it is a meeting. Schedule the assembly; keep the decision.
A report that includes everythingAn unfiltered recurring digest gets muted by week three, the same as any digest. Force 'material only' and a tight shape in the brief, or the Routine runs forever and nobody reads it.
Auto-writing before it earns trustA scheduled agent that edits the CRM unattended can do quiet damage on a cadence. Start every Routine read-and-report; let it propose CRM changes for a human until the proposals have been right for a while.
Never reading the outputThe value of a Routine is that a human acts on what it surfaces. If the Monday digest posts into a channel nobody opens, you automated the assembly and skipped the point. Route it where the reader already is.
Set-and-forget with no reviewMarkets and pipelines shift; a Routine tuned six months ago may be tracking the wrong things. Revisit each one when the underlying work changes, then leave it alone again.
Questions people ask
- How is a Routine different from just running the prompt myself each week?
- Only in one way, and it is the way that matters: it happens without you. Running the prompt by hand works right up until the busy week you forget, and busy weeks are exactly when the pipeline read is most useful. A Routine is a scheduled cloud agent, so the work shows up on time whether or not anyone remembered. You go from being the cron job to being the reader.
- What GTM work is actually a good fit for a Routine?
- Specifiable, repeatable work where the judgment is in reading the output, not producing it: a weekly pipeline-risk digest, a nightly new-lead brief, a competitor check, a monthly CRM-hygiene sweep. If you can write down exactly what the output should contain and where it goes, it is a Routine. If producing it needs a fresh human decision each time, it is not.
- Can a Routine write to my CRM, or only read and report?
- It can do both, but earn the write. Start every Routine read-and-report: it surfaces what it would change and a human acts. Once the 'would change' has been reliably right, let it write the safe things and keep a human on anything irreversible. A scheduled agent editing records unattended on day one is how a small rule error becomes a repeating one.
- Where do I learn how Routines actually work?
- The Claude Hub has a full guide at /claude/routines/, including how they differ from running something by hand and the pitfalls to avoid. This page is the GTM-specific version: which recurring plays are worth handing over, and how to brief them so the output gets read instead of muted.