1) The Three Axes of Prompt Design

• Structure: order, hierarchy, delimiters.
• Context: injected data, examples, role definitions.
• Intent: the real goal behind the words.

2) System → Instruction → Input Hierarchy

SYSTEM: You are a scientific writing assistant.
USER: Summarize this article in one paragraph.

Each layer narrows ambiguity. A good system prompt defines style and boundaries; the instruction sets task granularity.

3) Context Packing

• Use tables, JSON, or schemas — not prose — for structured data.
• Anchor facts before tasks (“Context first, command second”).
• Trim redundant adjectives; every token counts.

4) Few-Shot and Chain-of-Thought

Q: Translate "hello" to French
A: bonjour

Q: Translate "good night" to French
A: bonne nuit

Pattern repetition builds latent grammar in the model’s hidden states.

5) Anti-Patterns

• Vague adjectives (“creative”, “unique”) without examples.
• Nested or conflicting instructions.
• Excessive boilerplate or apologies — they dilute focus.

6) Prompt Chains for Agents

Break reasoning into atomic steps:

1. Rephrase the task.
2. Plan (list required sub-actions).
3. Execute step-by-step.
4. Summarize or verify output.

7) Evaluation Metrics

• Precision: fewer hallucinations.
• Consistency: same input → same output.
• Latency vs depth: response time vs token usage.

“Prompts are UX for the invisible.” — redesign.ir