Operating as artifact

Day-to-day operations (monitoring ads, watching user behavior, triaging support, watching social mentions) aren’t separate from the concept. They’re one more projection of it. The concept dictates how the thing behaves in the world over time; operations make that behavior real on an ongoing basis.

Most “AI in the business” conversations are about chatbots. The more interesting use is Claude as a recurring operator: running the small, repetitive tasks that keep the concept alive in production.

Monitoring ad accounts

Ads generate a daily stream of anomalies: creative that underperforms, targeting that starts overspending, a competitor entering an auction. Reviewing them manually is an hour a day; most teams don’t do it, and the spend silently drifts.

A Claude agent can:

• Pull yesterday’s spend, impressions, and conversion data from each platform
• Compare to a 7- or 28-day baseline
• Surface only the anomalies, the things worth a human glance
• Draft remediation options (pause this, increase budget here, test this new creative)

You review the summary in five minutes a day instead of missing the pattern for a week.

Monitoring user behavior

The same pattern applies to product analytics. A Claude session can run at a cadence and:

• Watch for feature-usage changes that might indicate a regression
• Flag cohorts whose engagement shifted (churn risk, activation drop)
• Surface quotes from user feedback that echo the data
• Connect it back to recent deploys to suggest cause

Good at: making you notice things you’d have missed. Not good at: replacing human judgment on what to do about them.

Monitoring social accounts and brand mentions

Claude can run scheduled scans across Reddit, X, Hacker News, and ProductHunt for mentions of your product, competitors, or relevant topics. Classify by sentiment. Flag engagement opportunities (real questions that could use a reply). Track competitor launches and pricing changes.

The output is a recurring scan report you can read in two minutes. The cost is one cron-scheduled Claude session. (Example: the social-media-monitor skill on this site.)

Support ticket triage

When a ticket arrives, Claude can:

• Classify it (bug, question, feature request, billing)
• Check if it matches known issues or recent deploys
• Draft a response for human review, not send it
• Escalate the hard ones with context

The human still reviews and sends. The time-per-ticket drops from ten minutes to one.

Routines: scheduling repeating tasks

Recurring tasks deserve first-class treatment. Three mechanisms, in increasing permanence:

• /loop: run a prompt or slash command on an interval from the current session. Good for during-work polling (“check the build every minute until it passes”) and lightweight day-of ops.
• Scheduled remote agents: triggers that fire on a cron schedule in Anthropic’s infrastructure, no local machine required. Good for standing routines: the morning ad scan, the weekly social summary, the nightly log triage. (See the schedule skill.)
• Routines baked into the application: the most durable form. Your product’s own scheduler runs the routine as ordinary code and calls Claude as a callable only for the steps that need judgment: classifying, summarizing, drafting. Fetching data, comparing to baselines, and recording results stay deterministic.

Routines can notify you: a summary pushed to Slack, a notification to your phone when something needs a decision. Run this way, an application watches itself: hourly health checks that record their findings, error detection that turns anomalies into enriched tickets, support triage that stages replies for one-tap approval. Faxterra runs on exactly this pattern.

When to keep a human in the loop

For anything that:

• Sends a message to a customer or publishes externally
• Spends money or changes pricing
• Creates tickets, PRs, or production changes
• Is irreversible

Draft-then-review is almost always the right pattern. The cost of a bad agent action with real-world impact dwarfs the minor inconvenience of a human confirmation step.

For internal monitoring, classification, and summarization (where the worst case is a missed signal), full automation is fine.

Resources

• The social-media-monitor and aeo-analysis skills on this site