ST: Prompting Strategies — Structure Note
A Structure Note (ST) is a lightweight MOC inside the Zettelkasten. It groups permanent notes on a single theme without duplicating their content.
Theme: How do different prompting strategies affect LLM performance on PM tasks?
Core Claims (linked permanent notes)
Cluster Summary
Six interconnected claims about how prompting strategy choice affects LLM performance and reproducibility in PM tasks.
- CoT improves structured classification → PN-CoT-FewShot-Prompting
- Few-shot k has non-monotonic effect on MAE → PN-CoT-FewShot-Prompting
- Local LLMs trade capability for reproducibility → PN-LLM-Local-vs-Cloud
- RCOIF structures AI prompts for research tasks → PN-RCOIF-Framework
- EGI maps unfamiliar domains iteratively → PN-EGI-Framework
- AMI enables iterative self-improvement → PN-AMI-DRCA-IIPR-Frameworks
PUMA Experiment Connection
The four prompting strategies evaluated in Stage 1 and Stage 2:
| Strategy | Code | Literature basis |
|---|---|---|
| Zero-Shot | S1 | Brown et al. 2020 |
| Few-Shot-3 | S2 | Tawosi et al. 2024 (CoGEE) |
| Few-Shot-6 | S3 | Calikli & Alhamed 2025 |
| Chain-of-Thought | S4 | Wei et al. 2022 |
→ EX-Hypotheses-H1-H2 → EX-Stages-Overview → PT-PUMA-Experiment-Prompts → LN-Calikli-2025-RequestFormats — non-monotonic k effect → LN-Datasets-JiraSR-TAWOS — evaluation datasets → PN-IssueTriage-StoryPoints — PUMA tasks measured → MOC-Methods-Frameworks — All prompting methods