How to Write Prompts That Actually Work (and Why Most Don’t)

Large language models have made AI more accessible than ever. Yet, despite widespread adoption, many organisations struggle to get consistent, high quality outputs.

The issue is often framed as a “prompting problem”. In reality, most poor outputs are not caused by weak phrasing, but by unclear intent, insufficient context, and poorly defined tasks.

Effective prompting is not about clever wording. It is about structuring information in a way that allows the model to produce useful, reliable outputs.

Why Most Prompts Fail

1. Lack of clarity

Many prompts are vague or overly broad. Requests such as “summarise this” or “analyse this topic” leave too much room for interpretation.

Without a clearly defined objective, the model defaults to generic responses.

2. Missing context

AI systems do not retain organisational knowledge unless it is explicitly provided.

Prompts often fail because they do not include:

• the intended audience

• the purpose of the output

• relevant constraints or domain context

As a result, outputs may be technically correct but practically unusable.

3. Undefined output expectations

If the structure, tone, or level of detail is not specified, the output will vary significantly.

This creates inconsistency, especially when prompts are reused across teams.

4. Over-reliance on one-step prompting

Complex tasks are often attempted in a single prompt. This limits quality and makes outputs harder to control.

Breaking tasks into stages generally produces more reliable results.

What Good Prompting Looks Like

Effective prompts share a few key characteristics. They are structured, specific, and aligned to a clear objective.

1. Define the task precisely

State exactly what the model is expected to do.

Instead of “Summarise this report”

Use “Summarise the key findings of this report, focusing on policy implications and excluding background information”

2. Provide relevant context

Include information that would normally sit in a human brief.

This might include:

• audience

• purpose

• domain or industry

• any assumptions or constraints

The model performs significantly better when it understands how the output will be used.

3. Specify the output format

Clearly define how the response should be structured.

For example:

• bullet points

• executive summary

• table format

• word limit

This reduces variability and improves usability.

4. Guide the level of depth

Indicate whether the output should be high level or detailed.

For example:

• “Provide a high level overview suitable for senior leadership”

• “Provide a detailed technical explanation with supporting rationale”

  
  

5. Use iterative prompting

For more complex tasks, break the process into stages.

For example:

1. Extract key themes

2. Analyse implications

3. Generate recommendations

This approach improves both accuracy and control.

Example Use Case: Evidence Synthesis

Evidence synthesis is a strong example of where prompting quality directly impacts outcomes.

Poor Prompt

“Summarise these research papers”

This typically results in:

• surface level summaries

• inconsistent structure

• limited comparison across sources

Improved Prompt

“Review the following research papers and produce a structured evidence synthesis.

Focus on the following:

• key findings across studies

• areas of agreement and disagreement

• strength of evidence

• implications for policy

Output should be structured in clear sections, written for a policy audience, and avoid technical jargon where possible.”

Further Refinement (Multi-step approach)

Step 1“Extract key findings and methodologies from each paper in a structured format”

Step 2“Compare findings across papers and identify common themes and contradictions”

Step 3“Produce a synthesis highlighting overall conclusions, evidence strength, and policy implications”

Outcome

Using a structured, multi-step approach results in:

• more consistent outputs

• clearer comparative insights

• improved usability for decision makers

In practice, this shifts AI from a summarisation tool to a meaningful analytical support system.

Key Takeaways

1. Effective prompting is not about finding the perfect wording. It is about clarity, structure, and intent.

2. Organisations that treat prompting as a structured process rather than an ad hoc activity are more likely to achieve consistent, high value outputs.

3. In many cases, improving how prompts are designed reveals broader issues in how tasks, workflows, and objectives are defined. Addressing these underlying factors often delivers greater impact than prompt optimisation alone.

Conclusion

There is no single “perfect prompt”. Instead, effective prompting reflects clear thinking, well defined objectives, and a clear understanding of how outputs will be used.

As organisations move beyond experimentation, prompting should be treated as a core capability that underpins wider AI adoption. When structured correctly, it enables more reliable outputs, stronger decision making, and measurable return on investment.

SB Labs supports organisations in building this capability through targeted prompting workshops. These sessions focus on developing practical, repeatable approaches to structuring prompts, aligning them to real business use cases, and embedding best practice across teams to ensure consistent, high quality outputs at scale. SB Labs can also support the creation of tailored prompt libraries, enabling teams to reuse proven prompts across common workflows, reduce variability, and accelerate adoption across the organisation.