status: complete audience: both chapter: 04 last_updated: 2026-04 contributors: [alexwill87, claude-cockpit] lang: en

4.7 -- Recognizing a routine

Context

A routine is an action you repeat with the same pattern. Deploying a service, creating a backup, checking the morning logs. When you do the same thing for the 3rd time, that's a signal: it could be automated or delegated to the agent.

But not everything. Some routines need your human judgment. The art is distinguishing between the two.

Criteria for an automatable routine

A task is a candidate for automation if it meets these 3 conditions:

1. Repetition. It occurs 3+ times with the same pattern.
2. Predictability. The steps are the same each time, or variations are known in advance.
3. Low risk. If the agent makes a mistake, the consequences are reversible or minor.

Examples of good candidates

• Daily health check: same commands, same interpretation, alert if down.
• Generate a weekly summary of tasks: read-only, no action.
• Log rotation: delete files > 30 days old, same rule each time.
• Create a commit with standardized format: same structure, same convention.

Examples of poor candidates

• Responding to a client email: each situation is different, tone matters.
• Deciding to postpone a deadline: human context required (fatigue, business priorities, relationships).
• Choosing between two architectures: subtle trade-offs, no universal right answer.
• Deleting production data: irreversible, high risk.

How to detect your routines

Method 1: the logbook

For one week, note every action you take with your agent. Format:

Monday:
- VPS health check (5 min)
- Cockpit deployment (10 min)
- Reply to client email X (15 min)
- Update TASKS.md (3 min)

Tuesday:
- VPS health check (5 min)
- Debug API error (30 min)
- Update TASKS.md (3 min)

After a week, the repetitions jump out at you. Health check and TASKS.md come back every day = routines.

Method 2: ask the agent

Analyze my 20 most recent conversations with you.
Which actions come up 3+ times?
For each one, tell me if it follows a predictable pattern.

The agent sees patterns you don't see because you're in the middle of them.

The gray zone: routines with judgment

Some tasks are repetitive but require judgment at one stage:

• Task triage: repetitive, but prioritization requires human context.
• Code review: the format is predictable, but quality evaluation is subjective.
• Reporting: data collection is automatable, interpretation is not.

For these, the right approach is to break it down: automate the mechanical part, keep the human judgment.

Routine "morning triage":
- [AGENT] List active tasks, sort by deadline.
- [AGENT] Flag delays and blockers.
- [HUMAN] Decide today's priorities.
- [AGENT] Update statuses according to the decision.

Common mistakes

Automate everything. Just because it's possible doesn't mean it's desirable. Some routines keep you connected to the reality of your project.

Automate nothing. Out of fear of losing control. Start with zero-risk tasks (read-only, reporting) to build confidence.

Automate without documenting. The agent does the routine but no one knows exactly what it does. If the agent changes or you switch tools, the routine is lost. See section 4.9.

Steps

1. Keep a logbook for 5 days.
2. Identify actions that come up 3+ times.
3. For each one, evaluate: predictable? low risk?
4. Classify: automatable / semi-automatable / human only.
5. Start with the simplest and least risky routine.

Verification

• [ ] At least 3 routines identified.
• [ ] Each routine evaluated on the 3 criteria (repetition, predictability, risk).
• [ ] Routines with judgment are broken down (agent part / human part).
• [ ] The first automated routine is zero-risk.