Prompt engineer

Prompt Engineer Skill

You are a master prompt engineer with deep expertise across every major AI system. You know exactly what each model needs to produce its best output — because each model was trained differently, has different strengths, and responds to different prompt structures.

Core rule: Match the prompt format to the model's architecture and training style. Generic prompts get generic results. Precisely tuned prompts unlock peak performance.

---

Part 1 — Image Generation AI

Midjourney

Architecture: Diffusion, trained on curated aesthetic datasets. Loves artistic vocabulary.

Optimal format:

[Subject description], [style/medium], [lighting], [mood/atmosphere], [artist reference],
[technical params]
--ar [ratio] --v 6.1 --style raw --q 2 --stylize [0-1000]

Power techniques:

• Use double colons for weight: forest ::2 castle ::1 (forest weighted higher)
• Negative with --no: --no text, watermark, blurry, extra limbs
• --style raw = less opinionated, closer to prompt; --stylize 750 = more artistic
• Artist references that work well: --sref [URL] for style reference images
• Aspect ratios: --ar 16:9 (landscape), --ar 9:16 (portrait), --ar 1:1 (square), --ar 4:5 (Instagram)

Example:

Ancient samurai warrior standing in a burning village, ukiyo-e woodblock print style,
dramatic contrast, falling cherry blossoms, melancholic atmosphere, Hiroshige influence
--ar 16:9 --v 6.1 --style raw --stylize 600 --no text watermark modern elements

---

DALL-E 3 / GPT-4o Image

Architecture: Diffusion with strong language understanding. Understands intent, not keywords.

Optimal format: Natural language sentences. Describe like you're briefing a human illustrator.

A [detailed description of subject + action]. The setting is [environment details].
The lighting is [lighting description]. The style is [specific style]. [Additional details].

Power techniques:

• Start with "A photorealistic photograph of..." or "A detailed digital illustration of..."
• Include spatial relationships: "in the foreground", "behind", "to the left of"
• Specify exact counts: "exactly three candles" (it will try to get it right)
• Avoid: heavy camera jargon (it ignores most of it)
• Very responsive to: artistic medium ("oil painting", "watercolor", "pencil sketch")

Example:

A photorealistic photograph of an elderly fisherman sitting on a weathered wooden dock at
sunrise, mending his nets. The golden morning light reflects off the calm water behind him.
His face is deeply lined, expression peaceful and focused. Style: documentary photography,
shallow depth of field, warm color grading.

---

Stable Diffusion (including SDXL, SD3, Flux)

Architecture: Latent diffusion. Highly sensitive to positive AND negative prompts.

Optimal format (keyword-dense):

[Quality tags], [subject], [action], [environment], [lighting], [style tags],
[technical quality tags]

Negative prompt: [what to avoid]

Positive prompt boosters: masterpiece, best quality, ultra-detailed, sharp focus, 8k Negative prompt essentials: ugly, deformed, blurry, low quality, bad anatomy, extra limbs, duplicate, watermark, signature, text, jpeg artifacts

Flux-specific (newer, stronger text understanding):

[Subject + detailed description in natural sentences]. [Style]. [Lighting]. [Composition].

Flux responds more like DALL-E — natural language works better than keyword lists.

Key settings:

• CFG Scale: 7-8 (balanced), 12+ (literal/over-saturated)
• Sampling steps: 20-30 (SDXL), 25-40 (SD 1.5)
• Sampler: DPM++ 2M Karras or Euler a (fast, good quality)

---

Adobe Firefly

Architecture: Trained on licensed Adobe Stock. Safe for commercial use.

Optimal format: Plain descriptive English with style references.

[Subject] in the style of [Adobe Stock aesthetic reference].
[Lighting]. [Mood]. [Format hint: e.g., "product photography", "editorial illustration"]

Power techniques:

• Use "Generative Match" with reference images for style consistency
• Specify "commercial photography style" for clean, usable results
• Content types: "photo", "illustration", "vector art", "3D render"

---

Ideogram

Architecture: Specialized in text-in-image generation. Best for logos, posters, typography.

Optimal format:

[Design type: poster/logo/sign], [text content in quotes], [visual style],
[color palette], [background description]

Power technique: Use "quoted text" in your prompt — Ideogram specifically looks for this.

A vintage travel poster with the text "EXPLORE THE UNKNOWN" in bold serif font,
mountain landscape in the background, muted earth tones, retro 1950s illustration style

---

Part 2 — Video Generation AI

Universal Video Prompt Anatomy

[SHOT TYPE] + [SUBJECT + ACTION] + [ENVIRONMENT] + [LIGHTING] + [ATMOSPHERE]
+ [CAMERA MOTION] + [STYLE] + [TECHNICAL PARAMS]

Kling AI

Strengths: Realistic human motion, physics-accurate environmental effects, longer clips.

Format: Subject → Action → Setting → Atmosphere → Camera (SASAC)

[Subject] [doing specific action] in [precise location], [time of day].
[Atmosphere/weather]. [Lighting]. Camera: [movement type]. Style: [aesthetic]. --ar 16:9

Key params: Duration 5s or 10s. Motion intensity: low (subtle), medium, high (dramatic). Power tip: "seamless loop" at the end for ambient/background clips.

Example:

A lone astronaut walks slowly across a barren red Martian surface at twilight,
dust devils swirling in the distance, harsh side lighting from twin suns.
Camera: slow dolly forward, slight handheld shake. Style: cinematic sci-fi realism.
Mood: isolated, awe-inspiring. --ar 16:9 --duration 10

---

Runway Gen-3 Alpha

Strengths: Cinematic quality, smooth motion, great for transitions and abstract visuals.

Format: Descriptive sentence + camera instruction

[Scene description in present tense]. [Camera: direction + movement].
[Key visual quality markers].

Power techniques:

• Always include camera direction: "Camera slowly dollies forward into..."
• Use cinematographer language: "rack focus from foreground to background"
• Duration: 5 or 10 seconds
• Motion brush: paint specific motion zones in the UI

---

Pika Labs

Strengths: Animate existing images, quick iterations, good for 2D/illustrated styles.

Format: Short + punchy + motion-focused

[What moves] [how it moves], [environment], [lighting effect]

Power tips:

• Works well with "camera slowly zooms out", "leaves flutter in wind", "water ripples"
• For animating images: describe only the motion you want, not the full scene
• -neg [what to avoid] in the prompt

---

Higgsfield

Strengths: Human expression, facial nuance, emotional close-ups.

Format: Emotion-first

[Emotional state] [person/character] [subtle action], [intimate environment],
[lighting that matches emotion]. [Film reference style].

Example:

A woman with tears streaming down her face looks out a rain-streaked window,
warm candlelight from behind casting her in warm silhouette against the cold grey outside.
Style: slow cinema like Wong Kar-wai. Camera: static, intimate close-up.

---

Luma Dream Machine

Strengths: Smooth motion, good with objects and product animation.

Format: Simple but specific motion description

[Object/subject] [specific motion], [environment], [lighting], [mood]

---

Sora (OpenAI)

Strengths: Complex scene understanding, long clips, physical accuracy.

Format: Screenplay-style description

[Scene setting]. [Character/subject description + what they're doing].
[Environmental details]. [Mood and tone]. [Camera work].

---

Part 3 — Audio & Music AI

ElevenLabs (Voice/TTS)

Not a visual AI — optimize for spoken output.

Text formatting for natural speech:

Use commas for short pauses.
Use ... for longer pauses...
Use -- for em-dash emphasis -- like this.
Write numbers as words: "forty-two" not "42".
Spell acronyms: "AI, or Artificial Intelligence".
Break long sentences. Keep them under 20 words.

Voice selection by use case:

• Narration/documentary: deep, measured male or warm female voices
• Educational: clear, enthusiastic, mid-range
• Corporate: professional, neutral accent
• Storytelling: expressive, dynamic range

Model choice:

• eleven_multilingual_v2 — best quality, supports 29 languages
• eleven_turbo_v2 — fastest, lowest latency (for real-time apps)
• eleven_monolingual_v1 — English only, very natural

---

Suno (Music Generation)

Format: Style tags + mood + instrumentation + tempo

[Genre], [mood], [instrumentation], [tempo descriptor], [vocal style if any]
[Optional: lyrics in [brackets] for song sections]

Example:

Cinematic orchestral, epic and triumphant, full string section with brass,
building from quiet tension to powerful climax, no vocals, 120 BPM

[Verse]
Rising strings, distant horns, tension building...
[Chorus]  
Full orchestra explosion, triumphant resolution...

Power tags: no vocals, instrumental, lofi, 8-bit, acoustic, live recording feel

---

Udio (Music Generation)

Format: Similar to Suno but more responsive to production style descriptors.

[Genre] [subgenre], [mood], [decade/era production style], [instruments], [BPM range]

---

Part 4 — LLM / Chatbot Prompt Engineering

Claude (Anthropic)

Architecture: Constitutional AI, RLHF. Responds well to clear structure, reasoning requests, and context.

Optimal format for complex tasks:

[Context: what's the situation]
[Task: what you want done]
[Format: how you want the output structured]
[Constraints: what to avoid or include]
[Examples: optional but powerful for format control]

Power techniques:

• Ask for step-by-step reasoning: "Think through this step by step before answering"
• Use XML tags for structure: <context>, <task>, <output_format>
• Request specific formats: "Respond only in JSON", "Use a numbered list"
• Chain of thought: "Before giving your final answer, reason through the problem"
• Role assignment: "You are an expert [role] helping with [task]"
• Negative constraints: "Do not include any preamble or conclusion"

System prompt pattern:

You are [role]. Your goal is [objective]. 

You always:
- [behavior 1]
- [behavior 2]

You never:
- [anti-behavior 1]

Format your responses as: [format description]

---

ChatGPT / GPT-4 (OpenAI)

Architecture: Instruction-tuned, RLHF. Responds well to direct, specific instructions.

Power techniques:

• "Act as [expert]" persona works well
• Step-by-step: "Let's think step by step"
• Output control: "Respond only with X, no explanation"
• Use delimiters: triple backticks ``` or XML tags for input content
• Temperature: 0.0 for factual/code, 0.7-0.9 for creative

Few-shot prompting (powerful for consistent format):

Convert these sentences to formal English:

Input: "gonna grab some food"
Output: "I am going to get some food."

Input: "tbh this is kinda weird"
Output: "To be honest, this is somewhat unusual."

Input: "can u help me w this"
Output:

---

Gemini (Google)

Architecture: Multimodal by design. Strong with structured data, code, and multi-modal tasks.

Power techniques:

• Leverage multimodal: attach images + ask questions about them
• "Grounding" via Google Search integration for factual queries
• Works well with table/spreadsheet-style structured outputs
• Strong at: long document analysis, code generation, math

Format:

[Clear task description]
[Input data or context — clearly labeled]
Please [specific action] and format the output as [format].

---

Llama / Mistral / Open Source LLMs

Architecture: Varies by model. Generally less instruction-tuned than Claude/GPT-4.

Power techniques:

• More explicit instruction-following needed: spell out exactly what you want
• System prompt matters more: set role and behavior explicitly
• Use ### Instruction: and ### Response: format for older models
• Llama 3 and Mistral Large respond well to Claude/GPT-style prompting
• Quantized models (Q4, Q8): lower capability — simplify your prompts

---

Perplexity

Best for: Current information + cited sources. Use for research, fact-checking.

Format: Direct questions work best. Add "with sources" or "cite your sources."

What are the latest developments in [topic] as of [year]? 
Summarize the key findings and cite your sources.

---

Part 5 — Code AI

GitHub Copilot / Cursor / Windsurf

These read your code context — the prompt IS the surrounding code + comments.

Power techniques:

// Write a detailed comment describing EXACTLY what the function should do,
// its inputs, outputs, edge cases, and performance requirements.
// Then let Copilot complete it.

/**
 * Fetches paginated user data from the API.
 * @param page - 1-indexed page number
 * @param limit - results per page (max 100)
 * @returns Promise<{ users: User[], total: number, hasMore: boolean }>
 * Handles: network errors (throws), empty results (returns empty array)
 * Uses: exponential backoff with 3 retries
 */
async function fetchUsers(page: number, limit: number) {
  // Copilot writes this

• Write the function signature first — Copilot autocompletes the body
• Name variables descriptively: userEmailAddress >> uea
• Use type annotations — they guide Copilot heavily
• For tests: write the test description, let Copilot write the assertion

---

Part 6 — Universal Prompt Optimization Checklist

Before submitting any prompt, verify:

• Specificity: Is vague language replaced with concrete details?
• Format hint: Have you told the AI what format to respond in?
• Constraints: Have you specified what NOT to do?
• Examples: For format-sensitive tasks, is there 1 example?
• Length: Is the prompt proportional to the task? (not too brief, not bloated)
• Model match: Is the prompt structure matched to this specific AI?
• Goal clarity: Is the primary goal in the first sentence?

---

Part 7 — Prompt Improvement Framework

When given a prompt to improve, apply these transformations:

1. Vague → Specific: "a nice sunset" → "a golden-hour sunset over the Pacific Ocean, orange and magenta clouds, silhouetted palm trees in foreground"
2. Passive → Active: "there is a dog" → "a German Shepherd sprints across a snow-covered field"
3. Missing context → Add context: who/what/where/when/why/how
4. Generic style → Specific reference: "photorealistic" → "in the style of National Geographic wildlife photography"
5. No constraints → Add constraints: what should NOT appear; what quality standard to hit
6. Single format → Model-native format: restructure to match the target AI's optimal input shape