Content moderation screens what users post or list. Social media platforms must block prohibited images and videos, and e-commerce platforms must block prohibited items. ChatGPT with careful prompt engineering can screen listings against a company's Standard Operating Procedure (SOP), which is the document that defines what is prohibited.

Workflow

The work runs as a cycle, so the SOP gets better each round:

graph TD
    A[Step 1: ChatGPT reviews the SOP] --> B[Step 2: ChatGPT classifies example products]
    B --> C[Step 3: ChatGPT gives step-by-step reasoning]
    C --> D[Step 4: Human reviews ChatGPT results]
    D --> E[Step 5: ChatGPT produces a refined SOP]
    E --> A

Sequence of prompts

The example category is 'Alcohol and Tobacco'. Each prompt has 1 key point:

StepPrompt doesKey point
1Provide input to define the objectiveAsk ChatGPT to summarise the key characteristics the SOP uses
2Provide a case and ask for analysisGive a catch-all answer of "Unclear" besides Yes and No
3Ask for step-by-step reasoningAsk ChatGPT to refer back to the SOP in its reasoning
4Ask for feedback and revisionUse the human-corrected labels to revise the SOP

The prompts themselves:

1. Read this Standard Operating Procedure (SOP) document which defines 'Alcohol and Tobacco'.
   [Insert the SOP text]. Please summarize the key characteristics that the SOP uses to
   define 'Alcohol and Tobacco'.
 
2. Here is a product description: [Insert product description]. Based on the SOP definition
   of 'Alcohol and Tobacco' you summarized, does this product fall under the category of
   'Alcohol and Tobacco'? Answer with 'Yes', 'No', or 'Unclear'.
 
3. Provide a step-by-step reasoning for your previous answer about the product
   classification, detailing how you used the SOP definition to determine if the product
   is 'Alcohol and Tobacco'.
 
4. Based on a series of product classifications (Yes, No, Unclear) that you previously
   provided and which have now been reviewed and corrected by a human expert, we need to
   update the SOP definitions to better align with these reviewed answers. Please revise
   the SOP to reflect the insights gained from the corrected classifications.

Why the key points matter:

  • "Unclear" stops the model from forcing a guess when the SOP does not settle the case
  • referring back to the SOP keeps the reasoning grounded in the rules, not in the model's general knowledge
  • the revision step is the feedback loop: the model's weights never change, only the SOP text does

Prompt engineering keys

KeyMeaning
Clarity of objectiveGive each prompt a clear, specific goal so the AI knows exactly what is expected
Sequential logicArrange prompts in a logical order that guides the AI through a thought process
Specificity and detailInclude details and specific instructions to reduce ambiguity
Feedback incorporationUse the AI's answers to refine the next prompts

Remembering tip

C S S F: "Clear Steps, Specific Feedback". Clarity, Sequential logic, Specificity, Feedback.

Brief prompt against detailed prompt

The same rule shows up when asking an LLM to extend a sentence with at least 3 object attributes:

  • a brief prompt only asks for 3 attributes, so the output format changes from sentence to sentence
  • a detailed prompt names the attribute types (counts of the object, color, action) and adds an example of the exact output format, so every answer follows the same layout

Prompts also give a chatbot its voice when it recommends products, as in the Recommendation Chatbot.