The course was one day on 15/01/2026.
| Strategy | What it is | When to use it | Why it works | Common failure mode |
|---|---|---|---|---|
| Zero-shot prompting | Ask directly with no examples | Simple, well-defined tasks | Leverages pretrained generalisation | Vague outputs if task is underspecified |
| Few-shot prompting | Provide a small number of examples | Patterned or stylistic tasks | Anchors the model’s behaviour | Overfitting to examples |
| Role / persona prompting | Assign a role or identity | Tone, domain framing, perspective | Narrows the solution space | Over-theatrical or inaccurate expertise |
| Instruction-first prompting | Clear task instruction before context | Complex tasksk | Sets priority and intent early | Buried or conflicting instructions |
| Delimiter-based prompting | Use markers to separate sections | Long or mixed inputs | Reduces context confusion | Inconsistent delimiters |
| Chain-of-thought (explicit) | Ask for reasoning steps | Logic, math, analysis | Forces structured reasoning | Overlong or wrong reasoning |
| Implicit reasoning | Ask for an answer only | When correctness > explanation | Avoids hallucinated reasoning | Harder to debug errors |
| Step decomposition | Break task into steps | Multi-stage workflows | Lowers cognitive load | Missing dependencies between steps |
| Self-critique / reflection | Ask the model to review its output | Quality-sensitive outputs | Catches inconsistencies | Rubber-stamping its own mistakes |
| Constraints & guardrails | Explicit rules and limits | Safety-critical or strict outputs | Shrinks the solution space | Conflicting constraints |
| Output schema prompting | Specify format or structure | Machine-readable outputs | Improves consistency | Fragile if schema is unclear |
| Tool-aware prompting | Instruct how/when to use tools | Hybrid LLM + system workflows | Separates reasoning from execution | Tool misuse or overuse |
| Temperature steering | Adjust randomness indirectly via phrasing | Creative vs precise tasks | Influences entropy | Not deterministic |
| Adversarial prompting | Probe edge cases or failures | Testing robustness | Exposes weak assumptions | Can induce hallucinations |
| Context compression | Summarise before reasoning | Very long inputs | Preserves signal, drops noise | Loss of critical detail |