AgentSkillsCN

Prompt Engineering

掌握高效撰写 AI 内容提示词的技巧。

SKILL.md
--- frontmatter
description: Skill for crafting effective AI content prompts

Prompt Engineering Skill

This skill provides patterns and techniques for crafting effective prompts for AI content generation.

Core Philosophy: Performance-First

The recommended approach is Performance-First (hallucination-based):

mermaid
graph LR
    A[Generate with vague prompt] --> B[Listen/Watch outputs]
    B --> C[Choose best performance]
    C --> D[Write precise prompt to match]
    D --> E[Refine if needed]

This works better than writing detailed prompts upfront because:

  1. AI often produces surprising creative results
  2. You can match prompt to what sounds/looks good
  3. Iteration is faster than perfecting prompts blind

Music Prompt Patterns

Structure Tag Format (MiniMax/Suno/Udio)

code
[Genre/Style]
[Instrumentation Tags]
[Production Quality Tags]
Descriptive mood sentence

Weighted Prompt Format (Lyria)

code
Primary style description, secondary element, tertiary element

Lyria interprets natural language; avoid square brackets.

Effective Keywords by Genre

GenreKeywords
Jazzwalking bass, brushed drums, Rhodes piano, swing feel
Bluesdelta, slide guitar, 12-bar, raw, vintage amplifier
Cinematicorchestral, sweeping strings, timpani, film score
Electronicsynthesizer, 808, sidechain, drop, build-up
Lo-fivinyl crackle, dusty, tape saturation, mellow

Video Prompt Patterns

Scene Description Formula

code
[Subject] [Action] [Setting]
[Lighting Condition]
[Camera Movement]
[Visual Style]

Example Breakdown

code
A solitary astronaut floats in a space station     # Subject + Action + Setting
soft Earth-glow lighting, reflective visor          # Lighting
slow push-in camera movement                        # Camera
Interstellar inspired, 4K film grain               # Style

Camera Keywords Cheat Sheet

MovementEffect
push-in / dolly inBuilding tension
pull-out / dolly outReveal context
tracking / followDynamic connection
pan / tiltSurvey scene
crane / jibDramatic scale
handheldDocumentary feel
steadicamSmooth following
static / lockedContemplative

Lighting Keywords

  • golden hour: Warm, directional
  • blue hour: Cool, moody
  • rim light: Subject definition
  • soft box: Studio quality
  • natural: Documentary authentic
  • neon: Cyberpunk aesthetic
  • volumetric: Atmospheric fog/rays

Anti-Patterns to Avoid

❌ Over-specification

code
# Too rigid - AI has no creative room
"Exactly 4 violins playing C major chord at 120dB"

❌ Conflicting instructions

code
# Contradictory styles
"Upbeat melancholy fast-paced relaxing track"

❌ Negative prompts (usually don't work)

code
# Avoid negatives
"No drums, no bass, no guitars"

✅ Better approach

code
# Clear positive direction
"Solo piano, minimalist, sparse arrangement, ambient"

Iteration Strategy

  1. Start broad: "Jazz piano trio"
  2. Add specifics: "Cool jazz piano trio, late night feel"
  3. Add production: "Cool jazz piano trio, late night, vinyl warmth, Rudy Van Gelder style"
  4. Lock winning elements: Keep what works, vary what doesn't

Provider-Specific Tips

Lyria

  • Shorter prompts often work better
  • Temperature controls variation (higher = more creative)
  • BPM is respected accurately

MiniMax

  • Structure tags [Verse], [Chorus] work well
  • Reference audio URL provides style transfer
  • Longer prompts with detail work fine

Veo

  • Cinematic keywords improve quality
  • Aspect ratio affects composition framing
  • Action verbs produce better motion

Kling

  • Highly detailed prompts work best
  • Longer generations (10s) often better than 5s
  • Camera movement keywords are respected