Give the agent
a finish line.
The first draft
The useful revision
"Build a single-file garden simulator with a 24 by 16 grid. Let me toggle cells, step once, run, pause and reset. Use the neighbor rules printed below. Save and restore the grid through validated JSON text. Check a stable block, a blinker and invalid input. Keep the page usable at 390 pixels wide."
The revision names the screen, simulation rules and controls, then gives observable checks. JSON is the structured text used to save the grid; validation checks that text before restoring it. A block stays still and a blinker alternates direction, giving the agent known examples to check.
Keep the instructions close to the task
A skill is a reusable set of task instructions an agent can load, such as a procedure for reviewing a program. OpenAI’s September guidance recommends focused skills with clear descriptions and relevant supporting detail. Accumulated instructions can become irrelevant or contradictory. [7]
For your next project, list the five behaviors that matter most. Add one example of success and one example of failure. Ask the agent to propose the simplest implementation that covers them.
Keep a short decision record. A rule such as "invalid saves preserve the current garden" gives the agent a concrete behavior to protect.