AI Pair Programming Workshop: Habits for Reliable Output

Classic pair programming has two roles. The driver types, the navigator thinks about direction and catches mistakes. With a coding agent, the agent often becomes the driver, and the human becomes the navigator. That sounds comfortable, but it demands discipline: a navigator who stops paying attention is no longer pairing, just watching. The workshop is built around habits that keep the human mind engaged.

Five habits that keep the human engaged

Habit 1: Think aloud, in writing

When you pair with a person, you explain your reasoning. With an agent, you do the same by writing a short note about intent before asking for code. Something like "I want to handle the empty case first because it is the one that crashed yesterday" gives the agent better context and clarifies your own thinking. In the workshop, participants write these notes for every task and notice how often the act of writing reveals a flaw in the plan.

Habit 2: Keep steps small

A human pair rarely writes five hundred lines before talking. An agent can, which is exactly the danger. We practice asking for one logical step at a time: a function, a test, a screen state. The smaller the step, the easier it is to verify, and the cheaper it is to discard.

Habit 3: Verify before continuing

• Run the test suite or the relevant single test.
• Open the app and use the feature, including one unhappy path.
• Read the diff, including files you did not expect to change.
• Ask the agent to explain any line you do not understand, then judge the answer yourself.

Habit 4: Switch roles on purpose

Sometimes the human should drive. Tricky algorithms, delicate refactors and design-heavy code often go faster when you write the core yourself and let the agent handle the surrounding parts: tests, edge case lists and documentation. In the workshop we run exercises where participants must decide in advance who drives which part, and then reflect on the result.

Habit 5: Notice drift

Long sessions accumulate confusion. The agent starts repeating earlier mistakes, ignores a convention or contradicts itself. We teach participants to notice these signs and respond: summarize the current state, start a fresh session with focused context, or step away from the agent. For long-lived guidance, a project instruction file helps; see AGENTS.md and CLAUDE.md.

What goes wrong without these habits

Teams that skip the habits tend to describe the same symptoms. Sessions feel fast but the results need rework. Developers cannot explain code they merged an hour ago. Bugs reappear because the agent fixed the symptom but not the cause. Reviewers are surprised by files they did not expect. None of these are failures of the tool in isolation. They are what happens when the navigator disengages. We show these patterns with short, realistic examples so participants can recognize them in their own work, and then practice the countermeasure for each, such as stating the expected diff size before the agent starts and checking afterwards whether reality matched.

Workshop structure

1. A short demo of a bad pairing session, where the human stops thinking, followed by a better one.
2. Paired exercises with human pairs and an agent: one participant narrates, the other types the prompts, then they swap.
3. A debugging exercise where the agent produces a plausible but wrong fix and the pair must find the actual cause.
4. A reflection round on what felt different from pairing with a human.

How we see accountability

Under our principle of AI speed, human responsibility, the agent is a very fast collaborator but not a colleague who shares liability. At apprat.io, senior AI Code Engineers plan the work, review every change, test on real devices and take accountability for the result. Pairing habits are how that accountability becomes visible in daily work. If you are building the routine across a team, our article on prompt-to-pull-request training shows how pairing connects to review and merging. To bring this workshop to your team, book a call for free.