Getting an Agent to Compare Its Options
My weigh-options-and-decide prompt asks an agent to compare real alternatives across the considerations that matter before recommending a choice. A small settings example shows what that comparison should reveal.
“Here are three options. I recommend the second.”
An answer like that can look considered even when the recommendation feels like a guess. The agent has named some alternatives, but it hasn't explained how they compare on the things that matter for this problem.
That's the habit my weigh-options-and-decide prompt addresses. It asks for the comparison before the recommendation: establish the goal, examine the real alternatives, and explain what each gains and gives up. Choosing comes after that work.
1. Give the comparison something to turn on
Take a hypothetical application that needs to remember a sidebar's width. The width must survive a page refresh, but it doesn't need to follow the user between devices. The application already has a browser-preference helper and an account-preference service.
The agent could keep the value in component state, use the browser helper, or save it through the account service. A useful comparison would look like this:
| Option | Survives a refresh? | Where the change belongs | Main cost in this example |
|---|---|---|---|
| Component state | No | Sidebar component | Fails a requirement |
| Existing browser helper | Yes, in that browser | Sidebar and its existing preference adapter | No cross-device synchronization |
| Account-preference service | Yes, across devices | Sidebar plus the account service | Adds a service dependency for this setting |
These are the stated properties of the invented application. In real work, the agent needs to inspect whether the helpers actually behave that way.
Here, the browser helper fits. Component state fails the refresh requirement. The account service supplies synchronization the task doesn't need, while making the setting depend on another service.
Now change one requirement: the width should follow the account between devices. The same comparison points toward the account service. The recommendation changes because the requirement changed, not because one approach is always better.
That is the useful part of the exercise. The answer reveals what the choice depends on.
2. Compare the relevant aspects, then choose
The full prompt also asks the agent to question a familiar default, consider alternatives nobody has mentioned, and include a combination or doing nothing when either makes sense. It doesn't require a giant matrix for every decision.
For the sidebar, refresh behavior and synchronization settle most of the choice. A decision about a shared database might instead depend on compatibility, query patterns, or the cost of changing existing callers. Applying the same checklist mechanically would bury those differences.
Nor does assigning “maintainability: 8/10” make the comparison more convincing. A concrete explanation of which callers change is easier to assess than an unexplained score.
This resembles a process reminder such as “think step by step,” with a specific job attached: compare the candidates using the actual constraints before naming a favorite. It isn't a guarantee of a better answer. A comparison built on a false assumption can still produce a confident recommendation, so its important claims need checking.
3. Use the prompt at a real decision
The Prompt Library page has the complete prompt to read or copy, plus the skill download. Once it's available to the agent, a short request can supply the task:
Use weigh-options-and-decide to choose how this sidebar should remember its width. It must survive a refresh; cross-device sync isn't needed. Inspect the existing browser and account preference helpers.
The prompt also tells the agent to carry out the chosen next step when that's part of the authorized task. That follow-through is useful, but the reason for this prompt is the comparison that earns the recommendation.
A good answer leaves the deciding factor visible. If a requirement or project fact turns out to be wrong, there's a clear place to reconsider the choice.