7 Principles for Using AI Coding Agents Well

Start with the core

The quality of an AI agent's output is proportional to the quality of your instructions. Before switching models, check these seven things — most problems disappear.

1. Define the task precisely

Not "fix login," but "fix the bug where email validation passes an empty string, and add a test for the failing case." State the outcome and the done-condition.

2. Give enough context, but not too much

Provide the relevant files, conventions, and constraints — without polluting the context with unrelated information. Hand broad exploration to a subagent and take back only the conclusion.

3. Split into small units

One decision at a time. Break large work into stages, each small enough to verify.

4. Build in verification

"Done" means "confirmed with a test," not "written." Require the agent to run and verify its own work.

5. Separate roles

Splitting exploration, implementation, and review across distinct perspectives raises quality. Adversarial review — asking an agent to refute its own result — is especially powerful.

6. Work reversibly

Commit often in small units, and do risky changes in an isolated workspace.

7. Make it observable

If you can't see what an agent is doing, you can't control it. Keep a board or log that tracks progress.

Let the tool enforce these

Marblo turns these principles into structure — define tasks as tickets, assign roles to agents, and track progress on a kanban board. Riding the flow beats trying to remember the principles.