Start from the final deliverable
Decide what you want to hold at the end: a decision brief, a diagram specification, a code review, or a first draft. Choose one skill that directly supports that deliverable. A business decision usually benefits from a comparison before it benefits from a polished speech about the recommendation.
Use the worked examples to understand the kind of work a skill asks for. The example content is prepared material, not a promise that a model will reproduce its exact layout or quality. In particular, an illustration in the library does not mean a text model will render an image.
Add a second skill for a specific reason
A comparison followed by a phased plan can work well: first weigh the options, then plan the chosen direction. A research brief followed by a visual explanation can also be useful if the visual is grounded in the evidence. State the relationship so the model knows what the second step should use.
If two skills ask for different audiences, incompatible formats, or opposite levels of detail, resolve that tension in the brief. You might ask for a detailed internal analysis and a separate short public summary. Naming two outputs is clearer than expecting one response to be both exhaustive and very short.
Make the order tell a story
Arrange the recipe in the order you would do the work. Establish context before comparing options, compare before recommending, and check evidence before polishing the wording. PromptForge lets you reorder selections before creating the prompt.
Read the compiled prompt once from beginning to end. Look for instructions that repeat without adding value, assumptions introduced too late, and an output format that no longer fits the task. The compiler assembles the instructions; your review establishes whether they make sense for this particular job.
Use a preset as a starting point
A curated recipe can reduce the work of getting started, especially when you know the outcome you need but not the available approaches. Review each selection and adapt the brief. A preset for a launch cannot know your capacity, audience, timing, or business constraints.
After you find a combination that works, save your own recipe preset. Keep a name that describes the job, such as “Compare two reversible options.” Avoid a name that promises a guaranteed result. A reusable approach is most useful when you can recognize when its assumptions no longer apply.
Improve one thing at a time
If an answer is vague, first add the missing context. If it is difficult to use, clarify the deliverable or format. If it invents facts, restrict the evidence and explicitly ask for unknowns. Adding another skill is not always the right repair.
Save a prompt before making a meaningful revision. Compare versions of the prompt as well as the responses, because model outputs can vary even for the same input. Treat a single impressive answer as something to inspect, then test whether the approach is useful on a second realistic task.
Every skill should do a distinct job in the recipe.