Prompt Engineering in Practice

Five Fixes That Make AI Output Actually Usable

Poor AI output quality is usually a prompt design problem, not a model capability problem. The five fixes below each address a different type of prompt deficiency, with before-and-after examples for each. For the broader context of why prompt design matters, see The Engineering Source of the Personalised AI Experience Gap.

Fix 1: Add Role Framing and Task Constraints

Before: "Write me a follow-up email."

The difference: role (partner), context (two-week-old meeting), format constraint (150 words), and prohibitions (sales language) are all specified simultaneously. The usability gap is not a percentage — it is the difference between usable and needing a complete rewrite.

Fix 2: Explicitly Declare Output Format

Before: "Summarise the main risk points in this contract."

Explicit format declaration means output is directly usable rather than requiring secondary reformatting. When AI output needs extensive restructuring before it can be used, the cause is almost always that the format requirement was never stated.

Fix 3: Require Step-by-Step Reasoning

For tasks that require multi-step logic, ask the AI to show its reasoning process rather than jump directly to conclusions.

The effect of step-by-step reasoning is dual: AI conclusions are generally more accurate, and the reasoning chain lets you verify each step rather than accepting conclusions blindly.

Fix 4: Use Negative Examples to Specify What to Avoid

Telling the AI what output types to avoid is often more effective than describing what you want.

Before: "Write an analysis of Hong Kong's current AI regulatory landscape."

Negative examples are most effective in scenarios where AI tends to produce templated, formulaic output.

Fix 5: Iterate Rather Than Chase the Perfect Single Prompt

The most effective prompting workflow is a series of progressive refinements, not the search for a prompt that succeeds on the first try.

In practice: first round generates a draft; second round: "Keep paragraphs one and three, rewrite paragraph two as a more concise version, remove paragraph four"; third round: "Add an opening sentence that signals the document's purpose to the reader." Each round's instruction is anchored to the previous round's output rather than re-describing the full requirement from scratch. Accumulated adjustments are more efficient than repeated resets.

Summary

The five fixes can be combined selectively by task type: role framing for vague task descriptions, format declarations when output structure matters, step-by-step for complex reasoning tasks, negative examples when AI output tends to be generic, and iterative refinement instead of searching for a perfect single prompt.

Further reading: Cursor vs Claude Code: Architecture and Use Cases