• 1. Zero-shot
• 2. Few-shot
• 3. Chain-of-Thought (CoT)
• 4. Self-Consistency
• 5. Tree of Thoughts (ToT)
• 6. ReAct
• 7. Reflexion

Prompt caching

Both Gemini and Anthropic support it. If the example block is identical across all 2M requests, you can cache it and pay a fraction of the input cost on cache hits (often ~10–25% of normal input price). This can drop the bill 60–80% with zero accuracy change and one config flag.

dynamic prompting

detect the complexity of the incoming query, and then you decide which technique to use in your prompt itself.

• dp methods

The technique ladder, cheapest → most expensive, looks roughly like this:
Each rung up costs more in latency, dollars, complexity, and maintenance.

1. zero-shot → few-shot → few-shot + CoT → RAG (retrieve definitions/examples) 
         → fine-tune → train your own
         
2. prompt caching → smaller model → retrieval few-shot (fewer examples per call) 
         → fine-tune
         
zero-shot direct       →  1x  cost
few-shot               → ~5x 
CoT (1 path)           → ~20x cost (longer output)
self-consistency N=10  → ~200x cost
ReAct                  → ~Nx 
Reflexion (3 tries)    → ~3x of whatever the base trajectory costs
ToT (50 nodes,eval ea) → ~2000–5000x cost
thinking               → variable

zero-shot           — no scaffolding, trust the model
few-shot            — show pattern by example
CoT                 — show your reasoning
self-consistency    — N paths, majority vote (fixes noise)
ToT                 — tree search with evaluation (rarely right in production)
ReAct               — reason + act + observe (the dominant agent pattern)
Reflexion           — self-critique + retry (fixes bias, needs evaluator)

• Definition gap → few-shot or system prompt clarification
• Knowledge gap → RAG, or fine-tune if the knowledge is stable + huge
• Reasoning gap → CoT, self-consistency, or ToT
• Format gap → strict output schema, few-shot showing the format, or constrained decoding