$ man context-wiki/agent-mode

Agent mode is the default execution mode. Claude reads files, writes code, runs terminal commands, creates new files, modifies existing ones, and builds features end to end. It is the mode where work actually gets done. Joe Rhew described it as a "teammate that needs onboarding." That is the right mental model. You would not throw a new hire into your GTM motion without context about your product, personas, and messaging. Same goes for agents. The quality of agent mode output depends entirely on the context you provide before execution starts.

PATTERN

Think of every agent mode session as onboarding a new teammate. What do they need to know? Your product (what you sell and to whom). Your voice (how you communicate). Your workflows (how things get done). Your constraints (what to avoid, what formats to follow). If you skip the onboarding, the agent will make assumptions. Some will be right. Many will not. And the worst part is that agent mode assumptions look confident. The output reads well. It just does not match your reality. Front-load the context. Let the agent execute from a position of knowledge, not guesswork.

PRO TIP

If you skipped plan mode on something complex, agent mode will charge ahead and make assumptions. It will pick an approach without evaluating alternatives. It will create files without checking if similar ones exist. It will write code that works but does not match your architecture. If you planned first, agent mode executes clean. It follows the plan. It creates the right files in the right locations. It uses the patterns you already established. The planning is not wasted time. It is the context that makes agent mode effective.

ANTI-PATTERN

Giving vague instructions and expecting specific output. "Make the site look better" is not actionable. "Change the hero section background to dark, increase heading font size to 3rem, and add 2rem padding" is actionable. Not loading relevant files before asking for changes. If you want Claude to match your existing code patterns, make sure it has read the existing code first. Letting agent mode run too long without checking output. Review after each major step. Course-correct early. A small misunderstanding in step 1 compounds into a big mess by step 10.