Prompting formula

The ReAct formula

Reason -> Act -> Observe loop; the foundation of agent prompts.

Skeleton

You may reason, then call a tool, then observe the result, repeating until done.

Agentic tasks with tools; research workflows.

Fill every slot. A formula with a missing slot silently degrades into a generic prompt — the model will fill the gap with its own assumptions.
Keep the ordering. Structure carries meaning; moving the output contract to the front of a long prompt is the most common reason a good formula fails.
Pin the output shape explicitly (format, length, delimiters) rather than describing it in prose.
Test with three inputs: a typical one, an edge case, and one where the correct answer is "insufficient context".

CO-STAR — Any LLM task where business context matters; default for the Master Engineer.
RTF — Quick, bounded requests; function-calling prompts.
CRISPE — Creative writing, strategy, brainstorming, marketing.
TAG — Crisp one-shot tasks; agent instructions.
Few-shot — Format-sensitive output, classification, style mimicry.
Chain-of-Thought — Math, logic, multi-step reasoning; improves accuracy on hard tasks.
Zero-shot CoT — Any hard question when no examples are available.
Self-Consistency — High-stakes reasoning; reduces variance.
Tree of Thoughts — Open-ended problem solving, planning.
Persona — Domain-expert output; tone control.
Output Format Contract — Programmatic output, extraction, structured data.
Chain-of-Verification — Fact-sensitive answers; hallucination reduction.
Step-Back — Hard problems where models rush to details.
Positive Instruction + Negative Boundary — Safety, style control, compliance.
System Prompt Architect — Any fine-tuned system prompt; the Master Engineer's signature output.