The Prompt Engineering That Still Works
Despite constant claims that “prompting is dead” (usually from people selling fine-tuning services), structured prompting remains the most cost-effective way to control model behaviour. What has changed is that the techniques that work in 2026 are different from those that worked in 2023.
What Is Now Obsolete
- Role-playing hacks (“pretend you are an expert…”) — modern models respond to direct instructions; personas add noise
- Emotional manipulation (“this is very important to my career”) — no longer affects response quality in frontier models
- Excessive chain-of-thought forcing — models like Claude 3.7 Sonnet have built-in extended thinking; you do not need to manually prompt for it
- Jailbreak techniques — frontier models in 2026 have robust alignment; jailbreaks that worked in 2023 are patched
What Works in 2026
- XML/JSON structured output — asking for structured output dramatically improves reliability for downstream parsing
- Few-shot examples in the format you need — showing 2–5 examples of the exact output format
- Explicit constraint enumeration — listing what to do AND what not to do
- System prompt separation — keeping persona/instructions in system prompt, tasks in user turn
- Minimal prompts with clear tools — for agentic tasks, a short goal + well-defined tools outperforms elaborate prompting
The New Frontier: Prompt Programmes
The most sophisticated prompting in 2026 is not individual prompts but prompt programmes: orchestrated sequences where the output of one prompt is the input of the next, with validation and branching. DSPy (Stanford) and similar frameworks compile these programmes automatically, optimising individual prompts for end-to-end performance.
System Prompt Security
As AI agents handle more sensitive tasks, system prompt injection attacks (malicious content in retrieved documents trying to override agent instructions) are a real threat. Prompt engineers now need to think about injection defence — structuring prompts so attacker-controlled content cannot override system instructions.