Prompt engineering has matured from a curiosity into a genuine skill — and the most effective prompts are rarely written from scratch. Here is a systematic approach to getting dramatically better results.
Why First Prompts Underperform
The first version of most prompts contains three common problems: vague instructions that require interpretation (the AI guesses your intent), missing context (what’s the use case, who’s the audience, what format do you need), and under-specified constraints (what should the output NOT include). The gap between a casual prompt and an optimised one often determines whether AI delivers acceptable or excellent results — the same model, with a better prompt, produces substantially better output.
The Iterative Method
The most reliable optimisation approach: generate a response, identify specifically what was wrong or missing, add that information as an explicit constraint, and repeat. Three rounds of iteration typically produces a stable, high-quality prompt. Keep a prompt library: when you find a prompt formulation that works, save it as a template. Prompt quality deteriorates if you cannot remember what made the last version better than the one before. Good prompts take time to develop; save them.
DSPy and Automated Prompt Optimisation
DSPy (Declarative Self-improving Python) is a framework for programmatically optimising prompts — defining what you want in terms of input/output examples, then using the framework to find the prompt instructions that best produce those outputs. This approach moves from manual prompt crafting to systematic optimisation using the LLM itself as a feedback signal. For teams running LLMs at scale, automated prompt optimisation can reduce cost and improve consistency significantly. For individual users, the manual iterative approach is more practical.
The Prompt Optimiser Pattern
Ask Claude or ChatGPT to improve your own prompt: “Here is a prompt I’ve been using: [prompt]. It produces good results but [specific issue]. Can you rewrite it to address [specific issue] while keeping [what works]? Give me three versions with different trade-offs.” Using the model itself to critique and improve prompts — meta-prompting — is one of the fastest ways to improve prompt quality without deep technical expertise.
