Tutorial ID | T-23 |
Section | AI & Analysis |
Title | Reviewing and editing AI-generated report sections |
Subtitle | Fix length, structure, references, and content issues in AI-generated report sections before they reach the report. |
Before You Begin
You have generated at least one AI report section for this case and have access to AI Prompt Hub.
You know the case's expected report structure and page-length range, if one has been agreed for this case type.
What You Will Accomplish
By the end of this tutorial, you will know how to prevent the most common problems in AI-generated report sections at the prompt stage, and how to correct them if they still appear after generation. Every section you review will have the right length, a consistent structure, traceable back-references, and no unsupported assumptions or repeated content.
Why Does this Tutorial Matter
A generated report section is only as useful as it is easy to review. The right length, short descriptive headings, a consistent structure, and a back-reference for every claim let attorneys and reviewers move through a report quickly and trust what they find.
Skip these checks, and a report that's too long buries key findings, one that's too short omits them, an inconsistent structure hides where information lives, and a claim without a back-reference has to be traced manually — all of which slow review and raise the risk that an unsupported or duplicated statement reaches the final report.
What to Do if Something Goes Wrong
The list below covers the most common problems in AI-generated report sections. Each entry follows the same pattern: what you'll notice, why it likely happened, a real example, and how to fix it.
Problem: A generated report is much longer or shorter than the case actually needs.
Likely cause: The prompt didn't specify a page count or range, so the AI produced a length it judged appropriate rather than one that matches the case.
Example: A straightforward single-provider case generates a twenty-page report that buries a few key findings in pages of routine detail.
Fix:
Before writing your prompt, decide whether the case needs a fixed page count or a range, based on how clear the scope is and how large the record volume is.
Open AI Prompt Hub and place the page-count instruction at the start of your prompt, for example “create a five-page report.”
For an uncertain scope or a large record volume, specify a range instead, for example “create a report of up to ten pages.”
Run the report and check that its length matches your instruction before reading the full content.
If the length isn't right, adjust the page count by two to three pages and run it again.
Save the working prompt once you find the right length, so you can reuse it for similar cases.
Problem: Paragraph headings in a generated report read like full sentences instead of short labels.
Likely cause: No heading-length instruction was included in the prompt, so the AI produced descriptive sentences rather than concise headings.
Example: A heading reads “The patient was seen for lower back pain following a workplace incident” instead of a short label like “Initial Back Pain Visit.”
Fix:
Open AI Prompt Hub and add an instruction such as “Headings should be no more than five words.”
Run the report and review each paragraph heading to confirm it reads as a short label, not a sentence.
If a heading is still too long, read the paragraph and rewrite the heading yourself using no more than five words that reflect its content.
Problem: A generated report is missing sections, has merged sections, or doesn't follow the structure reviewers expect.
Likely cause: The prompt used to generate the report was changed from the approved version, or the document set for the case is incomplete.
Example: A treatment timeline section that normally appears on its own is merged into the diagnosis section, and a reviewer can't find it where expected.
Fix:
Open AI Prompt Hub and review the prompt used to generate the report.
Confirm the section order, headings, and content expectations match the approved prompt for this case type.
If the prompt was changed, restore the approved version and run the report again.
If the prompt is unchanged, open Files and check that all expected document batches are present and complete.
If documents are missing, upload them and run the report again.
Review the regenerated report section by section against the expected structure to confirm nothing is skipped, merged, or missing content.
Problem: A long paragraph in the report has no back-references, so a statement can't be traced to its source without a manual search.
Likely cause: The prompt didn't specify where or how often back-references should appear, so the AI generated the content without links to the source documents.
Example: A paragraph covering a two-year treatment history has a single citation at the end, leaving a dozen individual claims impossible to trace on their own.
Fix:
Open AI Prompt Hub and add a back-reference instruction, such as requiring a source link after each key statement.
Run the report and identify any large paragraph that contains multiple statements without back-references between them.
For each unlinked statement, use Search or AI Chat to locate the source document rather than opening file batches one by one.
Open the identified document and confirm it's the correct source before using it as a reference.
Insert a back-reference link inline next to the statement, pointing to the relevant source document.
Once every statement in the paragraph has a link, review it as a whole to confirm no claim is left untraceable.
Problem: A statement in the report reads as an opinion, prediction, or recommendation rather than something the records actually state.
Likely cause: The prompt didn't restrict the AI to the provided documents, so it supplemented missing or limited information with an assumption.
Example: A report predicts the patient will need future surgery, when no source document makes that recommendation.
Fix:
Open AI Prompt Hub and add a restriction instructing the AI to base the report strictly on the provided documents, with no assumptions, opinions, predictions, or forecasts.
Run the report and check each statement against the question: is this supported by a specific document, or does it go beyond what the records say?
Flag any statement that reads as a prediction, a recommendation, or an unstated conclusion about causation.
For each flagged statement, use Files, Search, or AI Chat to check whether a supporting document exists.
If a source exists, add a back-reference to the statement and remove the flag.
If no source exists, remove the statement or rewrite it to reflect only what the records state.
Problem: A key finding gets a brief mention while a routine event takes up a disproportionate amount of the report.
Likely cause: The prompt didn't include prioritization instructions, so the AI treated all events with roughly equal weight instead of reflecting their significance to the case.
Example: A single sentence covers a documented surgical intervention, while several paragraphs describe routine follow-up visits.
Fix:
Open AI Prompt Hub and add an instruction to focus on significant findings, such as primary diagnoses, surgical interventions, and documented changes in condition.
If you know which findings matter most to the legal argument, name them explicitly in the prompt.
Run the report and evaluate whether the space given to each event reflects its actual significance to the case.
Flag any section where a critical finding is underrepresented or a minor event is overrepresented.
For each flagged section, confirm the correct significance using the source documents, Search, or AI Chat.
Expand underrepresented sections with detail and back-references, and condense or remove sections that overstate minor events.
Problem: The same fact, diagnosis, or event appears more than once across different sections of the report.
Likely cause: Different source documents described the same event from different angles, and the AI treated each mention as a separate entry instead of recognizing them as one occurrence.
Example: A single hospitalization is described once from a physician's note and again from a discharge summary, so the report covers the same stay twice.
Fix:
Open AI Prompt Hub and add an instruction not to repeat information across sections.
Run the report and identify each repeated statement.
For each one, use Files, Search, or AI Chat to confirm whether the entries describe the same event rather than two separate occurrences.
Keep the most complete and well-referenced version of the statement and remove the duplicates.
Review the report as a whole to confirm each fact and event appears only once.
Tip Save any prompt instruction that fixes a recurring problem — length, headings, back-references, prioritization — as part of your standard AI Prompt Hub template for this case type, so future reports start closer to right the first time. |
Summary
Every generated report section is now checked for length, heading clarity, structure, traceable back-references, and unsupported or repeated content — with each fix built into the prompt itself so future reports start closer to right the first time.