One-Shotting

Also known as

One-shotting, "do it all in one prompt".

Context

You ask to "build an app" and expect a finished result after a single pass without intermediate checks. A successful demo can reinforce this expectation.

Problem

The first result looks complete, even though the agent has not yet verified the scenarios and constraints. The problem arises when this result is accepted as a finished application.

Why people do it

  • A demo shows a successful run and may not reveal how many attempts failed.
  • The agent quickly produces a convincing main scenario.
  • Planning and repeated checks seem unnecessary after a successful first result.
  • The context window seems infinite until it runs out.

Consequences

  • ➖ The window can run out in the middle of several unfinished parts.
  • ➖ A mistake in an early decision spreads into the code that follows.
  • ➖ The first run reveals defects that were invisible on reading.
  • ➖ The next session spends time figuring out what state the work is in.

Signs

  • The prompt describes a large release, and no intermediate results are planned.
  • The agent does not run the tests or the application along the way.
  • Several scenarios remain "almost working".
  • The next session starts with archaeology.

A better way

Use a quick first pass to explore an idea. For a production implementation, prepare Tracer-Bullet Tickets or a Feature List, finish one feature at a time, and verify each result with a Feedback Loop. Then a broken-off session leaves already verified parts and one current task.

Example

Before

Build a task tracker with teams, a kanban board, notifications, access permissions, and a dark theme.

After

Break the task tracker into tickets. The first one: create a task and see it on the board

results matching ""

    No results matching ""