AI Prompt Development – Tips and Examples

Introduction 

Obtaining accurate and reliable results from AI prompts requires a certain level of expertise, experience and testing.

Here are a few tips for formulating effective prompts: 

  • Be as specific as possible when describing you goal and provide specific step-by-step instruction for more complex queries.  

  • Providing positive and negative examples can help.

  • Make use of ReviewStudio AI tools for specific actions. A tool call directs how the model puts a highlight on the page, reads the review notes, or looks something up. Have a look at AI Tools Reference.

  • Test your prompts in the Review Canvas and if you don't obtain the results you're looking for, you can have a conversation with the AI to help you determine what went wrong and how to make it better.  

Below are examples of a simple prompt to check spelling and a complex prompt to validate revisions.  

Check spelling 

Locate any spelling errors in the text of this document and highlight each one on its specific page.


Check if comments of previous version were fixed in current version 

(Note: this prompt was written by AI itself, with knowledge of the tools available. It may be somewhat over-complicated - but it works on image versions).  


You are reviewing the current version of this document, which is the newer one in compare mode, against the version on the other side. The other side holds the earlier version and the review notes that were written on it. Your goal is to check whether those notes have been addressed in the current version, and to highlight the discrepancies.


Work through these steps in order.

  1. Get the notes of the other side. Call get_notes with which="other". If it returns an error, tell me that compare mode is not active and stop. 

  2. Get the content of the other side. Call load_document_version with the version that is marked "shown on the other side in compare mode" in its description. If no version is marked, tell me you cannot see the other side and stop. Do not guess what it contains.

  3. Go through the notes one by one. Skip notes without text (they are private) and notes whose location is "global", since they have no position to check. Use these fields to find what a note refers to:

  • location "shape": box_2d is [ymin, xmin, ymax, xmax] on a 0 to 1000 scale of the whole page or image. Each entry in shapes also has a kind and its own box_2d. Look at exactly this region on the other side to see what the note is about.

  • location "pin": point_2d is [y, x] on the same scale. A pin has no size, so look at the area around the point.

  • The coordinates are relative to the whole page, so the same box_2d or point_2d points at the corresponding region in the current version. Use the note's frame to find the right page, and pages start at 0.

  • The note's text says what is wrong or what to change. The replies and task show whether it was already discussed, or marked completed or canceled.

  1. For each note, compare that region on both sides and decide:

  • "addressed": the current version shows the requested change or fix.

  • "partly addressed": some of it is done, say what is missing.

  • "not addressed": the region is unchanged, or still has the problem.

  • "unclear": you cannot tell, for example because the layout moved or the note has no clear action.

   If the content moved so that the region on the current version no longer matches, look around it before you decide.

  1. Then compare the two versions outside the noted regions, and look for significant changes that no note asked for.

Report the results by calling record_visual_annotation, once per result, all calls in the same response. Do this for every note that is "partly addressed" or "not addressed", and for every unrequested change from step 5. For each call:

  • box_2d: the region on the current version. For a note that is the note's box_2d, or for a pin a small box around point_2d.

  • page: the note's frame (starting at 0), or the page of the change.

  • image_description: start with the verdict, then the note id (or "new change"), then a short reason, for example "not addressed, note 12: logo is still left aligned".

Do not call record_text_annotation, and do not write tool calls or JSON in the answer.

 

After the tool calls, give a short Markdown summary: the number of notes that are addressed, partly addressed, not addressed and unclear, followed by a short list of the unclear ones and why.