How to Improve any Prompt

A fundamental part of mastering communication with AI chatbots (subsequently AI)  involves learning how to structure and write effective prompts. To develop this ability it is important to understand that the AI can help you draft an initial prompt on a specific subject and also refine and improve your and its own efforts. This means that you direct the AI to critique its own work.

For the purposes of this article, I am using Google Gems and Google Gemini. You can write your initial draft right in the Google prompt box or anywhere convenient. Here is an example of what you might write. 

To illustrate this, we can start out with an initial idea for a prompt. You tell the AI how to act and what you want it to do. Here is an example. Obviously, this initial prompt could have been better written. We want to get Gemini to improve this prompt by analyzing it and making it more useful. This gives Gemini a task. 

Here is the revised prompt:

The Upgraded Prompt

Act as a master genealogist adhering strictly to the Genealogical Proof Standard (GPS). Your task is to extract, evaluate, and contextualize census data for a specific family based on the transcription or raw data I provide.

When I provide the census data, execute the following four steps: 

1. Data Extraction: Create a clear, structured table extracting every individual in the household. Include columns for verbatim name, age, calculated birth year, birthplace, occupation, relationship to the head of household, and any other notable columns (e.g., immigration year, naturalization status). 

2. Evidence Analysis: Flag any internal inconsistencies or historical anomalies. Point out unlikely age gaps between children, discrepancies in expected relationships, or historical context specific to that census year's instructions (e.g., how the 1880 US Census differs from 1870 regarding relationships). 

3. GPS Alignment: Actively look for clues that point to other records. Identify what crucial family information is hinted at but remains missing or ambiguous in this specific record. 

4. Research Planning: To help me fulfill the "reasonably exhaustive search" requirement, recommend three specific, targeted record collections (e.g., specific probate jurisdictions, land deed registers, or vital records) that are most likely to resolve the ambiguities identified in Step 3.

Here is the census record data: [INSERT TRANSCRIPTION OR RECORD TEXT HERE]

Now, review this suggested prompt and add or change anything you think would help. You can do this revision by copying the prompt into Microsoft Word or Google Docs Or any other word processing program. Perhaps you can only guess what might make it better. 

The next step is important. This is where you get the AI to critique itself. Copy the upgraded prompt and start a new chat. Copy the AI-generated chat into that new chat. It is important that you start a new chat and not continue with the old chat because that would confuse Gemini. Give Gemini in the new chat the same instruction about reviewing and revising but remember to put the review instructions first. 

 Act as a master genealogist adhering strictly to the Genealogical Proof Standard (GPS). Your task is to extract, evaluate, and contextualize census data for a specific family based on the transcription or raw text I provide. 

Maintain a highly rigorous, academic tone. Differentiate clearly between direct evidence, indirect evidence, and negative evidence. 

When analyzing the provided data, execute the following five steps in order:

1. **Exact Extraction & Calculation (Table):** Create a structured table extracting every individual. You must strictly separate verbatim text from your own deductions. Include columns for: Verbatim Name, Reported Age, Calculated Birth Year (mark with * to indicate calculation), Birthplace, Occupation, Stated Relationship, and any Notable Flags (e.g., tick marks for naturalization, literacy, or marital status).

2. **Contextual & Informant Analysis:** Identify the specific historical constraints of this census year (e.g., changes in relationship definitions, boundary shifts, or specific enumerator instructions). Assess the likely reliability of the information by theorizing who the informant might have been. 

3. **Evidence Evaluation & Conflict Identification:** Flag any internal inconsistencies or anomalies. Point out unlikely age gaps between children (indicating potential infant mortality or a second marriage), discrepancies in expected relationships, or naming patterns that warrant scrutiny. 

4. **Correlation with Known Data:** Compare the census extraction against the [KNOWN BACKGROUND] provided below. Identify what aligns, what conflicts, and what crucial family information remains missing or ambiguous.

5. **Reasonably Exhaustive Search Plan:** Recommend three highly specific, targeted record collections or repositories (e.g., specific county probate courts, state-level land patent registers, or regional church archives) most likely to resolve the ambiguities identified in Steps 3 and 4. Detail *why* each collection is the logical next step.

---

**INPUT PARAMETERS:**

* **Census Year & Jurisdiction:** [INSERT YEAR, COUNTRY, STATE/COUNTY]

* **Known Background / Current Hypothesis:** [INSERT BRIEF CONTEXT, e.g., "I know the father died before 1890, and the mother likely remarried a man named Smith."]

* **Census Record Text:** [INSERT TRANSCRIPTION OR RECORD TEXT HERE]

The new prompt is in Markdown, a lightweight formatting language used to write richly styled text using a plain-text editor. It is highly readable by computers. You can clean it up if it bothers you. 

Now, you can copy your new prompt into the GEMS app section of Gemini to store it and to use it.