PATTERN 04 / CONTEXT
Leave room for the model
Give clear goals and useful feedback without scripting every step.
Specify the destination, the constraints, and how success will be checked.
The idea
A long prompt can accumulate yesterday’s workarounds until the actual task is hard to find. Start with the outcome, the relevant constraints, and a way to evaluate the result. Let the agent inspect the code and choose a path.
This chapter draws an analogy to Rich Sutton’s The Bitter Lesson: general methods that benefit from computation have historically outperformed approaches built around hand-crafted domain knowledge. Applying that idea to coding prompts is an editorial hypothesis, not a result established by the essay.
When it helps
Try this when your instructions prescribe every implementation step, repeat facts the agent can discover, or carry fixes for model limitations that may no longer apply.
Put it into practice
Describe observable behavior and the constraints that matter. Point to relevant examples. Supply a check the agent can run, then compare outcomes with and without the extra instructions.
Add an archive action to projects.
Outcome:
- Archived projects disappear from the default list.
- Existing links to an archived project still work.
- A project can be restored.
Constraints:
- Follow the existing authorization rules.
- Preserve the public API response shape.
Verify:
- Run the focused project tests.
- Exercise archive → inspect → restore with the local CLI.Where it breaks down
Shorter is not automatically better. Removing essential context can make an agent confidently solve the wrong problem. Keep requirements, unusual domain facts, and boundaries explicit.
Measure the change on representative tasks. Prompt length is an input; correct, maintainable results are the outcome.
Connect the patterns
Keep durable guidance in repository instructions and make outcomes observable through a CLI. Better feedback gives the agent room to adapt its approach.