Prompt Design
The Difference is in the prompt
By The 1938 Group | 15 June 2026 | Johannesburg
Same tool, different outcome.
A Vast majority of prompts are written by people in the same way they text a colleague: informal, rushed, and with little thought. 'Summarise this.' 'Write an email.' 'Give me ideas for a presentation.'
Afterwards, surprise is received when the output is confidently wrong, off-point or too vague to be useful. AI is far from the finished article and isn't as good as advertised. The tool is supposedly at fault.
In truth the prompt is the culprit, not the AI tool.
A prompt should not be treated like a search query. It is a set of instructions being given to a powerful, context-blind system that knows nothing about your organisation, your client, your standards, or your professional obligations. Like everything in business and life, the quality of your inputs will determine the quality of your outputs.
The 1938Group's AI Usage & Data Safety Enablement Programme: prompt design. In this article we cover the RCTCF framework, the four safety rules that apply to every prompt, and the complete Prompt Library (40 tested templates across 11 business functions) offered by The 1938 Group, built with safety constraints embedded from the start.
Why Your Prompt is at Fault
A weak prompt can be identified fails in one of four ways. The best possible fix can be determined once the failure mode that the prompt falls into is identified.
• Too vague: "Summarise this report." The AI has no guidance on what matters, who the audience is, how long the summary should be, or what not to include. The output will reflect that vagueness back.
• No constraints: Without explicit boundaries, AI will fill in the gaps. It will invent statistics to correct the statement. Ai will infer beyond the source material. It will confidently fabricate details, also referred to as AI Hallucinations.
• Wrong/dangerous context: Providing too little context in the prompt will produce a generic, unhelpful outputs. Add too much context, especially context that includes RED or AMBER information, creates a data exposure problem. The context field of your prompt is the highest-risk component.
• No format specification: The AI will choose an output format if one is not specified. It may not select the needed format. By specifying the output structure in the initial prompt, will save on editing time and reduces the chances of content being in an unusable format.
THE CORE TRUTH | AI does not know what you need unless you tell it. Precisely, completely, and with constraints that protect the quality and safety. |
|---|
The RCTCF Framework: Five Components of a Safe, Effective Prompt
The 1938 Group's AI Usage Policy uses the RCTCF framework for prompt design across all business functions. Here is a full breakdown of what each component does, why it matters, and where safety implications sit:
| COMPONENT | THE QUESTION IT ANSWERS | EXAMPLE | WHY IT MATTERS |
|---|---|---|---|---|
R | ROLE | Who should the AI act as? | "You are a financial analyst specialising in SME reporting..." | Sets expertise level and perspective. Dramatically affects output quality for specialist tasks. |
C | CONTEXT | What background does the AI need? | "This is for a mid-sized SA retail company preparing a board update..." | Check every word here against RAG. RED data in context = data breach. |
T | TASK | What exactly should the AI do? | "Summarise sections 4-7, focusing on the three main risks identified..." | Be precise. Vague task = vague output. Specify scope, angle, and purpose. |
C | CONSTRAINT | What rules must the output follow? | "Do not invent statistics. Flag any uncertain claims. Maximum 200 words..." | Your primary prevention against hallucinations. Always include. |
F | FORMAT | How should the output be structured? | "Present as a numbered list, one sentence per item, with a bold heading..." | Specify when format matters for use. Saves significant editing time. |
A prompt doesn’t need to have all five components to be considered as passable. A simple reformatting task may need only T (Task) and F (Format). Complex research starting point may need all five. The principle is to use as many components as the task and risk level warrant, never skip C (Constraints) on any output you plan to use professionally.
Seeing the Difference: Before and After RCTCF
The gap between a weak prompt and a structured one is not academic. Here it is in practice, across three common business tasks:
Email drafting
✗ WEAK PROMPT | ✓ RCTCF PROMPT |
"Write an email about a project delay." | "You are a professional communication specialist. I need to write an email to a client (generic, do not use names). Situation: a technology implementation project will be delayed approximately two weeks due to technical integration challenges. Draft an email that: acknowledges the delay, explains the reason briefly, provides a revised timeline, and reassures on quality. Tone: professional, apologetic but confident. Length: 150-200 words. Do not include specific names, dates, or amounts, use [PLACEHOLDERS]. Do not make commitments beyond standard service terms." |
The second prompt produces a professional, usable draft. The first produces something you will spend more time editing than writing from scratch.
Research starting point
✗ WEAK PROMPT | ✓ RCTCF PROMPT |
"Tell me about POPIA compliance." | "You are a compliance research analyst. Provide an overview of South Africa's POPIA requirements for organisations that process personal data of employees and clients. Include: key principles, data subject rights, responsible party obligations, and consequences of non-compliance. Constraints: focus on general principles only, do not provide specific legal advice; indicate areas where current legal counsel should be sought; flag any areas where regulations may have changed recently. Format: structured overview with clear headings, maximum 400 words." |
The second prompt bounds the output, reduces hallucination risk on a legally sensitive topic, and explicitly directs the user to seek qualified advice where needed. That constraint is not bureaucratic. It is the difference between useful research scaffolding and a liability.
Meeting agenda
✗ WEAK PROMPT | ✓ RCTCF PROMPT |
"Create a meeting agenda for a project review." | "You are a meeting facilitation expert. Meeting type: project status review. Duration: 60 minutes. Attendees: project team of 5 plus a sponsor (generic, no names). Primary objectives: review progress, address blockers, confirm next steps. Create an agenda that: structures time for all workstreams to report, allocates problem-solving time for blockers, and confirms decisions and actions. Constraints: total time must not exceed 60 minutes; include a 5-minute buffer; format must be professional and clear. Format: timed agenda with section headings and owners indicated by [ROLE]." |
The 1938Group Prompt Library: 40 Templates, 11 Functions
Good prompts should not be rebuilt from scratch every time. The 1938Group Prompt Library provides 40 tested templates across 11 different business functions. Each prompt was built with the RCTCF structure, data classification notes, and verification guidance embedded into them.
FUNCTION | TEMPLATES AVAILABLE | RISK LEVEL |
|---|---|---|
General Business Writing | Document drafting, rewriting, summarisation, outlines | Low |
Communication | Email drafting, complaint responses, meeting requests, follow-ups | Low–Medium |
Analysis & Research | Research starting points, pro/con analysis, SWOT, comparative eval | Low–Medium |
Meeting & Documentation | Agendas, meeting summary templates, process docs, report structure | Low |
Human Resources | Job descriptions, interview frameworks, feedback structures | Medium |
Marketing & Communications | Content ideas, social media, email marketing, press releases | Medium |
Technical/IT | Code documentation, troubleshooting guides, requirements structure | Low–Medium |
Project Management | Project charters, status reports, risk frameworks, lessons learned | Low–Medium |
Customer Service | Response templates, complaint frameworks, FAQs, announcements | Medium |
Legal & Compliance | Structure only, ALL content requires qualified legal review | HIGH |
Finance | Structure only, ALL content requires qualified finance review | HIGH |
Two sections of the library carry explicit warnings that deserve emphasis here:
Legal and Compliance (Section 10)
AI at its current state shouldn’t be used to generate legal advice, legal documents, or compliance interpretations without a qualified legal review. The templates in this section provide structure only, outlines, document frameworks, compliance organisation guides. Every substantive legal element must be drafted or reviewed by a qualified legal professional. The fact that an AI output reads like a legal clause does not make it legally sound. This is a non-negotiable.
As a South African professional: AI cannot interpret POPIA, the Companies Act, or sector-specific regulations on your behalf. It can help you organise information and prepare questions for your legal counsel. It cannot replace them.
Finance (Section 11)
AI must not be used for financial analysis, calculations, or advice without qualified professional review. Financial decisions must be based on verified data and professional judgment. AI-generated financial figures, even when they look precise and authoritative, are not a substitute for finance team input. The templates here provide document structure only.
The Four Safety Rules That Apply to Every Prompt
The RCTCF framework governs structure. These four rules govern safety. They apply to every prompt, across every template, in every business function:
✓ | NEVER include RED-classified data in any prompt. Personal data, client confidential information, financial records, legal documents, health data, security credentials, and trade secrets must never enter a public AI tool. |
✓ | SANITISE AMBER data before inclusion. Use anonymisation, generalisation, hypothetical scenarios, aggregation, or selective extraction before any AMBER information enters a prompt. Apply the false anonymisation test. |
✓ | VERIFY all outputs before use. A well-structured prompt reduces the risk of hallucination; it cannot be eliminated. The VERIFY framework applies regardless of how good your prompt was constructed. |
✓ | YOU are responsible for all your AI-assisted work. The prompt library gives you structure and safety constraints. The professional accountability for the output sits with you, not the template, nor the tool. |
These rules are not suggestions. They are the foundation of responsible AI use, and they interact directly with the organisation's legal obligations under POPIA, professional confidentiality obligations, and the accountability framework established in the hallucinations article.
Prompt Library in Practice
The library is a starting point, not a script. Here is the workflow for using it effectively:
Select: Find the template that matches your task. If nothing matches exactly, find the closest one and adapt it.
Classify: Before you customise, classify all the information you plan to include. RED = stop. AMBER = sanitise first. GREEN = proceed.
Customise: Replace the [BRACKETED PLACEHOLDERS] with your specifics: keeping the safety constraints and RCTCF structure intact.
Check: Before submitting, confirm: no RED data included, AMBER data sanitised or approved, task clearly specified, constraints in place.
Verify: Apply the VERIFY framework to the output before using it. The template reduces hallucination risk. It does not eliminate it.
Contribute: If you develop a prompt that works exceptionally well for a recurring task, submit it for inclusion in the library. Effective prompts are institutional knowledge.
ON CONTRIBUTION | The best prompt libraries grow through use. When your team finds prompts that consistently produce high-quality, safe outputs for specific tasks, those prompts should be documented and shared. Prompt design is a professional skill. |
|---|
Prompt Design Is a Competitive Skill
At this point, most organisations are at the stage where AI tools are being used but prompt quality remains inconsistent. Some people are getting excellent outputs. Most are getting mediocre ones. The difference is almost entirely in the design of the prompt, which is teachable.
Teams that invest in prompt literacy, who understand the RCTCF framework, who use a consistent prompt library. Teams that embed safety constraints as habit rather than treating them as afterthoughts will consistently get better outputs, experience fewer errors, and create fewer data exposure incidents than teams that don’t.
This isn’t a marginal advantage; it is far more significant. Over the course of a year, across an organisation, the gap between disciplined prompt design and casual prompting compounds into a significant difference in output quality, risk exposure, and the time teams spend correcting AI errors instead of using AI outputs.
Follow 1938Group on LinkedIn for the weekly module updates or visit info@the1938group.co.za to book for the full workshop.
https://www.linkedin.com/company/the1938group/