Learning Objective: Tell AI which supplied or approved sources to use, how to handle missing support, and how to distinguish evidence from assumptions.
Lesson Purpose
Tell AI which supplied or approved sources to use, how to handle missing support, and how to distinguish evidence from assumptions. This lesson is part of the AI in a Pinch™ prompt-development pathway. The learner is not being taught to chase clever wording. The goal is to convert approved business intelligence into a prompt that is specific, reviewable, reusable, proportionate to the task, and safe to test.
A prompt is one component of a larger business workflow. It does not contain the whole business, replace source systems, authorize data use, approve claims, or transfer accountability to an AI provider. The person using the prompt remains responsible for selecting the task, protecting information, supplying accurate context, defining boundaries, reviewing the output, and deciding whether the result may be used.
Why This Matters
Poor prompts often hide a deeper problem: missing business facts, unclear customer understanding, an undefined offer, conflicting objectives, weak evidence, vague output expectations, or no human decision boundary. Adding more words does not solve those problems. The strongest prompt contains the smallest approved set of information that adequately directs the task.
This lesson also prepares the learner to use the V-Prompt Score™ as an improvement tool. The score rewards relevant completeness across seven dimensions, while the separate safety, evidence, and human-oversight gate prevents a high numerical score from being mistaken for authorization.
Business Scenario
A prompt asks for a factual report but does not supply sources or instruct the model to label uncertainty, leading to invented statistics and citations.
The responsible response is to identify what the prompt expects the AI system to guess, retrieve the approved inputs, clarify the task, define the output, add constraints, and name the reviewer before testing.
Core Concepts
Approved sources. Identify the documents, data, policies, notes, or links the task may rely on.
Source priority. State which source governs when information conflicts.
Evidence boundary. Do not permit unsupported facts, testimonials, statistics, or guarantees.
Uncertainty handling. Require the model to flag missing evidence, assumptions, and questions.
Citation expectation. Specify when citations, source labels, or traceable references are required.
The V-Prompt Architecture™
A practical V-Prompt contains the following layers:
- Purpose and task. State the immediate action AI may support and the business outcome it contributes to.
- Business context. Add relevant approved facts about the business, operating situation, and source of truth.
- Customer or audience context. Identify the primary audience, relationship stage, needs, awareness, and relevant concerns.
- Product or service context. Explain the active offer, scope, delivery, value, proof, limitations, and status.
- Direction and output. Define the action, sequence, deliverable, structure, length, required elements, and success conditions.
- Tone and next step. State the brand voice, situation-specific tone, language level, and appropriate call to action.
- Boundaries and review. State prohibited information, claims, scope, professional limits, uncertainty handling, human review, and stop conditions.
Not every prompt needs the same amount of detail. A short internal brainstorming task may require a lighter foundation than a customer-facing, financial, employment, safety, legal, or public communication. The level of context and control should be proportionate to the consequence of error.
Step-by-Step Method
- List approved sources.
- Set source priority.
- Prohibit unsupported claims.
- Require uncertainty labels.
- Specify citation format.
- Add human verification.
After completing the steps, read the prompt as if you were a reviewer who did not participate in its creation. Ask whether another responsible person could identify the task, sources, assumptions, restrictions, output, and approval requirements without relying on unwritten knowledge.
Detailed Business Application
Start with the approved handoff from AIP-01 and AIP-02. Retrieve the Business Foundation Brief™, Customer Clarity Profile™, Offer Clarity Brief™, proof register, brand voice rules, AI goal, and desired output. Do not rewrite those assets from memory when an approved version exists.
Select only the information relevant to the current task. Context should reduce uncertainty. Irrelevant details increase prompt noise, create contradictions, expose unnecessary information, and make version control difficult. Use labelled blocks where the prompt contains several kinds of context.
Separate facts from instructions. A fact describes the business, customer, offer, source, or condition. An instruction tells the model what action to perform. A constraint limits the action or output. A review instruction explains what the model should flag and what a person must approve. Mixing these categories can make the prompt harder to interpret.
Use explicit uncertainty handling. Direct the model not to invent missing facts, prices, dates, statistics, testimonials, sources, qualifications, availability, customer history, or business commitments. It should identify gaps, label assumptions, or ask a limited clarification question when necessary.
Keep human oversight operational. Naming a reviewer is not enough. The reviewer needs the source information, subject knowledge, time, criteria, authority, and ability to reject or escalate the output. The prompt can remind the system that the response is a draft for review, but that statement does not replace the actual review process.
Common Mistakes
- Telling AI to research without source rules
- Accepting invented citations
- Treating internal notes as public authority
- Removing limitations
- Using stale sources
Risk, Privacy, and Human Oversight Checkpoint
Before testing, confirm that the selected tool, account, data, and use are approved. Use fictional, anonymized, public, or otherwise approved non-sensitive information during learning. Do not enter passwords, authentication codes, personal customer or employee information, confidential contracts, financial account information, legal or health records, trade secrets, unpublished strategy, or proprietary methods into a public or unapproved AI tool.
A strong prompt score does not authorize a risky task. Stop or escalate when the prompt affects legal rights, employment, finance, insurance, health, safety, security, vulnerable people, regulated services, sensitive profiling, or another consequential decision without qualified oversight.
Practical Activity
Complete Source and Evidence Instruction Block™ in the learner workbook.
Create source-use rules for one prompt.
Record the approved inputs, source owner, prompt version, V-Prompt score where applicable, safety-gate decision, reviewer, and next action.
Reflection Questions
- What information does this prompt currently require the AI system to guess?
- Which input is supported by an approved source?
- Which instruction could be interpreted in more than one way?
- What information or claim must the prompt prohibit?
- Who can meaningfully review the output and what criteria will they use?
- What is the smallest useful improvement before testing?
Key Takeaways
- Prompt quality begins with approved business inputs.
- Relevance and clarity matter more than prompt length.
- The V-Prompt Score™ identifies readiness gaps but does not guarantee output accuracy.
- The safety, evidence, and human-oversight gate is mandatory.
- Tested prompts still require output review and approval in AIP-04.
Worksheet or Resource
- Source and Evidence Instruction Block™
- File: `AIP-03_AI_in_a_Pinch_Prompt_Building_and_V_Prompt_Quality_Learner_Workbook.docx`