Reducing your unknowns is the trainable skill of agentic coding — and the agent itself is how you train it

Thariq Shihipar names the actual skill underneath “good prompting”: “the best agentic coders have relatively few unknowns” in what they hand the agent — expertise shows up as synchronization between the prompt and both the codebase and the model’s behavior, not as verbosity or precision for its own sake. The reassuring half of the claim is that this is learnable: “reducing and planning for your unknowns is the skill of agentic coding… this is a skill you can improve at, by working with Claude.” The agent isn’t just the thing you’re directing — working with it is the training loop that improves your own ability to anticipate what you don’t yet know.

This gives The map is not the territory — with a capable-enough agent, finding your unknowns becomes the real bottleneck a concrete unit of progress: since the map will never be complete, the thing that actually improves session-over-session isn’t the plan, it’s your calibration for where your own blind spots tend to be. It’s a specific instance of the general pattern in The model already knows the answer — it just can't reach it without the right tool — just as a tool can unlock capability the model already had, using the agent as an unknown-detector unlocks a skill (calibrated self-awareness of your own gaps) that reading documentation alone wouldn’t train nearly as fast.