Claude Code Routines for Marketing Teams

Anthropic released Claude Code Routines in 2026 Q2. One slash command, /schedule, now spins up a managed Claude Code session with a cron-style or GitHub-event trigger, a connector list, and a steerable web view. The cron-and-webhook wrapper most marketing teams wrote on top of headless Claude Code in 2025 is now first-class infra.

For marketing operators the news is bigger than for developers. Marketing workflows are mostly scheduled or event-driven by nature. Campaign deploys, dashboard regressions, content drift, lead routing, attribution sanity checks. A routine is the smallest unit of an agent which reacts to a calendar or an event instead of waiting for a human to press enter.

TL;DR

Claude Code Routines collapses the 2 to 5 weeks of cron-job, auth, webhook, and Slack-notifier work into one slash command. Marketing teams who built the wrapper by hand lose the differentiator. The new strategic question is which marketing jobs deserve to react to events without a human in the loop. Three decisions define every routine: trigger, context, steerability. Six marketing workflows fit the shape today.

Key Takeaways

• Routines is a Claude Code feature, shipped 2026 Q2, behind the /schedule slash command
• Triggers come in two shapes: time-based schedule and event-based (native GitHub or POST webhook)
• Sessions run on Anthropic-managed infra with web, CLI, and desktop steering
• Anthropic reports 200% weekly PR growth on Claude Code since January 2026, absorbed by a single content owner using routines
• 3 decisions per routine: trigger, context, steerability
• 6 marketing workflows fit routines on day one: deploy verifier, content drift, MQL triage, competitor watch, dashboard drift, inbound question router

What does Anthropic mean by routines?

A routine is a Claude Code session under Anthropic-managed infrastructure with 4 inputs. The prompt defines the work. The connected repos give read and write access. The connectors expose tools like GitHub, Slack, Drive, and any MCP server. The trigger fires the session.

Triggers split two ways. Schedule triggers run on cron-style cadences (once a week at 10am, daily at 6am, hourly). Event triggers fire on a GitHub event (issue open, PR merge, release cut) or a POST request to a webhook with the payload passed in as context.

Anthropic handles 4 things the marketing team used to handle: hosting, session state, connector auth, and observability. The session stays interactive. Open it in the web view, ask a follow-up mid-run, push it in a different direction, resume it next week.

Action item: Run /schedule once inside Claude Code with any throwaway prompt. Walk through the follow-up questions Anthropic’s agent asks. The questions reveal the design constraints Anthropic baked in.

Why does this matter for marketing teams?

Marketing teams hit a specific tax when a Claude Code workflow leaves the laptop. Hosting the headless session somewhere reliable. Wiring secrets. Wrapping it in cron. Building a webhook receiver. Adding a Slack notifier. Logging the runs. Adding retry logic when the model 503s. Onboarding the second user.

The build cost runs 2 to 5 engineering weeks per agent for a senior platform engineer. Most marketing orgs do not have a senior platform engineer sitting idle. The work falls on the most technical person in MOps, who absorbs the build cost as a side project, or the agent stalls at proof of concept.

Anthropic reports weekly PR volume on Claude Code grew 200% since January 2026. The same Claude Code and Agent SDK docs surface was held by a single owner under this load. Routines moved the owner from sustaining maintenance to shipping new content.

The same shape applies in marketing. A single lifecycle marketer watching 12 nurture flows. A single SEO operator monitoring 80 high-intent pages for content drift. A single demand-gen lead reviewing the weekly lead-scoring queue. The fixed cost per workflow drops to a slash command.

Action item: Pick one Claude Code workflow your team runs manually today. Count the steps a human performs: opening the laptop, running the command, copying the output, pasting it to Slack. Each step is a routine candidate.

What are the three decisions for every routine?

Anthropic’s design surface for a routine compresses to 3 decisions. Trigger, context, steerability. The same 3 decisions apply to every marketing use case.

Decision 1. Trigger

Two options. Schedule or event.

Schedule fits work tied to a calendar. The weekly content drift review. The Monday-morning lead scoring queue. The end-of-month attribution refresh.

Event fits work tied to a thing happening somewhere else. A campaign deploy completing. An asset moving to a launch-ready state. A dashboard regression crossing a threshold. A new question landing in a designated Slack channel. A POST request from any system you control.

Schedule is easier to reason about. Event has more reach when the trigger is a clean signal and the response is mechanical.

Decision 2. Context

Context is the ceiling on how good the output gets. Claude only knows what you connect.

For most marketing routines context comes from 3 sources. The relevant code or content repos for read and write. The data sources via connectors (CRM, BI tool, ad platform). The notification destination (Slack, email, an issue tracker).

The mistake operators make here is under-connecting. Hooking Claude to the CRM but not to Slack means the routine works and nobody finds out. Hooking Claude to GA4 but not to the content repo means the agent sees the data and writes nothing.

Decision 3. Steerability

Steerability is how you keep the routine honest. Three patterns work.

Generator-critique. One routine produces an output (a PR, a summary, a Slack post). A second routine triggers on the first one’s output and reviews it. Each routine is small. The review catches the failure modes the producer misses.

Live mid-run intervention. Open the routine’s session in the Claude Code web view. Watch a run. Ask a follow-up question. Tell the agent to stop or change direction. The intervention model is the routine’s primary safety net against duplicate work or off-rails behavior.

Output verification. For documentation, render the page. For lead scoring, sanity-check the top 10 scored leads against last quarter’s closed-won. For campaign deploys, hit the URL and confirm the title tag is correct. The verification step is a third routine or a human review.

Action item: For one workflow your team runs, write out trigger, context, and steerability on one page. If steerability collapses to “we will trust the agent,” design a generator-critique pair before going to production.

Which marketing workflows fit routines first?

Six workflows map cleanly to the trigger and context model today. Three deserve detail. Three get a one-liner.

Campaign deploy verifier

Trigger. Webhook from the CD pipeline after a campaign page or template deploys.

Context. The site repo. A monitoring connector (Datadog, GA4, the BI tool watching homepage CVR). Slack for alerts. Optional: a Twilio connector for high-priority pages.

Steerability. Claude runs the investigation and posts a go or no-go recommendation in Slack. A second routine triggers on the recommendation and reviews the reasoning before any rollback. Humans approve the rollback for the first 4 weeks. After 4 weeks of clean decisions, the team turns on auto-rollback for low-traffic pages.

Sample prompt. Investigate the latest deploy. Compare the home and pricing pages against the last 7 days of CVR. Open a Slack thread in #releases with go or no-go, evidence, and a draft rollback command if no-go.

Content drift detector

Trigger. Weekly schedule, Mondays 7am local.

Context. The blog repo. The brand voice repo (the same context/voice.md and context/anti-ai-writing.md files yfxmarketer uses). The proof bank. A list of the top 50 ranking pages from Search Console (via the GSC MCP server). Slack for the report.

Steerability. Claude opens a PR in the blog repo with proposed updates to drifted pages. Human review on the PR. A second routine triggers on PR open and runs the anti-AI validator against the proposed text.

Sample prompt. Review the top 50 ranking pages against the brand voice rules and the proof bank. Open a PR for any page with more than 2 voice violations or stale data older than 6 months. Cite the rule and the fix in the PR description.

MQL backlog triage

Trigger. Weekly schedule, Mondays 9am local.

Context. CRM connector (Salesforce, HubSpot). Slack channel for the report. A scoring rubric stored in the marketing-ops repo.

Steerability. Claude posts the top 20 MQLs of the week with score, reasoning, and a recommended next play. A demand-gen manager reviews. Decisions feed back into the rubric. Over 8 weeks the rubric becomes deterministic enough to wire scoring upstream.

Three more routines worth wiring

• Competitor release watcher. Daily schedule. RSS or GitHub watch list. Posts a summary to Slack.
• Dashboard drift detector. Webhook from the BI tool when a metric crosses a threshold. Opens a Linear issue.
• Inbound question router. Event-triggered by a new question in #ask-marketing Slack or a new tag on a Zendesk ticket. The routine checks the brand content repo and the FAQ database, drafts a response, opens a review item for the field-marketing owner.

Action item: Pick one routine from this list. Write the trigger, context, and steerability spec on one page. Send it to the most senior engineer on your team for a 30-minute review before you run /schedule.

How do you steer a routine without breaking it?

The steerability question splits into 3 sub-questions.

How do you watch a run? Open the Claude Code web view at claude.ai under Code, then Routines on the left side panel. Click into a session. The session reads as if the routine were a normal Claude Code conversation. The web view supports stopping a routine mid-run when a redundant or wrong-direction action is in progress.

How do you keep the routine in scope? Two techniques. First, write the prompt with a hard out: If you do not find a documentation gap, post a one-line update to Slack and stop. Second, scope the connector permissions. A routine with read-only GitHub does no damage if the prompt drifts.

How do you catch silent failure? Generator-critique is the most reliable pattern. The producer routine ships a PR. The critique routine reads the PR and the source data and grades it against a rubric. The critique catches the failure mode of the producer becoming overconfident or stale.

Verify outputs as the third gate. For docs, render the page. For lead scoring, eyeball the top 10 against last quarter’s closed-won. For campaign deploys, hit the URL. Verification is cheap when the routine produces structured artifacts (PRs, issues, JSON).

Action item: For every routine you ship, name the steerability pattern on the routine itself. Generator-critique, live intervention, or output verification. If the answer is “none,” do not turn it on.

What changes about the marketing-ops operating model?

The first-order effect is the build cost dropping. The second-order effect is the higher-altitude question taking over.

The build-cost effect. A marketing team without a senior platform engineer now ships routines at the speed of the slash command. The 2 to 5 weeks per workflow drop to 30 minutes. The 6 routines listed above used to require a full quarter of platform work for an org of any size. Routines compresses the same scope to a sprint.

The selection effect. The new bottleneck is selection. Routines fits work which is repetitive, well-scoped, and has a clean steerability gate. Routines does not fit work which requires open-ended judgment, multi-stakeholder approval, or political nuance. Picking the work earns the payoff.

Director and VP read. The operating-model implication is sharper than the workflow implication. A marketing org running 12 routines under a generator-critique pattern moves the headcount conversation from “how many specialists do we hire” to “how many routines do we operate and who reviews the critique outputs.” Roles shift from execution to selection and oversight.

The risk. Random acts of agentic AI multiply when the build cost drops. Routines makes shipping easy. Routines does not make selection easy. The discipline lives in the rubric the team uses to decide which work earns a routine.

Action item: Map the 5 most repetitive workflows your team runs this quarter. For each, write trigger, context, steerability, and a one-sentence reason the work deserves to react to an event instead of a human.

Final Takeaways

Claude Code Routines collapses the cron-and-webhook tax to a slash command. The build cost per agent drops from 2 to 5 weeks to one prompt.

The 3-decision frame holds. Trigger, context, steerability. Every routine has all three or it does not run.

Steerability is the load-bearing decision. Generator-critique, live intervention, output verification. Pick one before you ship.

The marketing operating model shifts up the stack. Selection beats execution. The new question is which work earns a routine, not how to schedule it.

The discipline holds the gain. Routines makes shipping cheap. Random acts of agentic AI now ship faster than ever. The rubric for selection is what keeps the team out of an agent sprawl.