prompt-engineering

belt CLI スキルをインストール： npx skills add belt-sh/cli

プロンプトエンジニアリングガイド

inference.sh CLI を使って、AI モデル向けのプロンプトエンジニアリングを習得しましょう。

クイックスタート

inference.sh CLI（belt）が必要です。インストール手順

belt login

# Well-structured LLM prompt
belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "You are a senior software engineer. Review this code for security vulnerabilities:\n\n```python\nuser_input = request.args.get(\"query\")\nresult = db.execute(f\"SELECT * FROM users WHERE name = {user_input}\")\n```\n\nProvide specific issues and fixes."
}'

LLM のプロンプト作成

基本構成

[Role/Context] + [Task] + [Constraints] + [Output Format]

役割の指定

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "You are an expert data scientist with 15 years of experience in machine learning. Explain gradient descent to a beginner, using simple analogies."
}'

タスクの明確化

# Bad: vague
"Help me with my code"

# Good: specific
"Debug this Python function that should return the sum of even numbers from a list, but returns 0 for all inputs:

def sum_evens(numbers):
    total = 0
    for n in numbers:
        if n % 2 == 0:
            total += n
        return total

Identify the bug and provide the corrected code."

思考の連鎖

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Solve this step by step:\n\nA store sells apples for $2 each and oranges for $3 each. If someone buys 5 fruits and spends $12, how many of each fruit did they buy?\n\nThink through this step by step before giving the final answer."
}'

少数ショットの例

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Convert these sentences to formal business English:\n\nExample 1:\nInput: gonna send u the report tmrw\nOutput: I will send you the report tomorrow.\n\nExample 2:\nInput: cant make the meeting, something came up\nOutput: I apologize, but I will be unable to attend the meeting due to an unforeseen circumstance.\n\nNow convert:\nInput: hey can we push the deadline back a bit?"
}'

出力形式の指定

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Analyze the sentiment of these customer reviews. Return a JSON array with objects containing \"text\", \"sentiment\" (positive/negative/neutral), and \"confidence\" (0-1).\n\nReviews:\n1. \"Great product, fast shipping!\"\n2. \"Meh, its okay I guess\"\n3. \"Worst purchase ever, total waste of money\"\n\nReturn only valid JSON, no explanation."
}'

制約の設定

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Summarize this article in exactly 3 bullet points. Each bullet must be under 20 words. Focus only on actionable insights, not background information.\n\n[article text]"
}'

画像生成プロンプト

基本構成

[Subject] + [Style] + [Composition] + [Lighting] + [Technical]

被写体の説明

# Bad: vague
"a cat"

# Good: specific
belt app run falai/flux-dev --input '{
  "prompt": "A fluffy orange tabby cat with green eyes, sitting on a vintage leather armchair"
}'

スタイルキーワード

belt app run falai/flux-dev --input '{
  "prompt": "Portrait photograph of a woman, shot on Kodak Portra 400 film, soft natural lighting, shallow depth of field, nostalgic mood, analog photography aesthetic"
}'

構図の制御

belt app run falai/flux-dev --input '{
  "prompt": "Wide establishing shot of a cyberpunk city skyline at night, rule of thirds composition, neon signs in foreground, towering skyscrapers in background, rain-slicked streets"
}'

品質キーワード

photorealistic, 8K, ultra detailed, sharp focus, professional,
masterpiece, high quality, best quality, intricate details

ネガティブプロンプト

belt app run falai/flux-dev --input '{
  "prompt": "Professional headshot portrait, clean background",
  "negative_prompt": "blurry, distorted, extra limbs, watermark, text, low quality, cartoon, anime"
}'

動画プロンプト

基本構成

[Shot Type] + [Subject] + [Action] + [Setting] + [Style]

カメラの動き

belt app run google/veo-3-1-fast --input '{
  "prompt": "Slow tracking shot following a woman walking through a sunlit forest, golden hour lighting, shallow depth of field, cinematic, 4K"
}'

動作の説明

belt app run google/veo-3-1-fast --input '{
  "prompt": "Close-up of hands kneading bread dough on a wooden surface, flour dust floating in morning light, slow motion, cozy baking aesthetic"
}'

時間に関するキーワード

slow motion, timelapse, real-time, smooth motion,
continuous shot, quick cuts, frozen moment

高度なテクニック

システムプロンプト

belt app run openrouter/claude-sonnet-45 --input '{
  "system": "You are a helpful coding assistant. Always provide code with comments. If you are unsure about something, say so rather than guessing.",
  "prompt": "Write a Python function to validate email addresses using regex."
}'

構造化出力

belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Extract information from this text and return as JSON:\n\n\"John Smith, CEO of TechCorp, announced yesterday that the company raised $50 million in Series B funding. The round was led by Venture Partners.\"\n\nSchema:\n{\n  \"person\": string,\n  \"title\": string,\n  \"company\": string,\n  \"event\": string,\n  \"amount\": string,\n  \"investor\": string\n}"
}'

反復的な改善

# Start broad
belt app run falai/flux-dev --input '{
  "prompt": "A castle on a hill"
}'

# Add specifics
belt app run falai/flux-dev --input '{
  "prompt": "A medieval stone castle on a grassy hill"
}'

# Add style
belt app run falai/flux-dev --input '{
  "prompt": "A medieval stone castle on a grassy hill, dramatic sunset sky, fantasy art style, epic composition"
}'

# Add technical
belt app run falai/flux-dev --input '{
  "prompt": "A medieval stone castle on a grassy hill, dramatic sunset sky, fantasy art style by Greg Rutkowski, epic composition, 8K, highly detailed"
}'

複数ターンの推論

# First: analyze
belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Analyze this business problem: Our e-commerce site has a 70% cart abandonment rate. List potential causes."
}'

# Second: prioritize
belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "Given these causes of cart abandonment: [previous output], rank them by likely impact and ease of fixing. Format as a priority matrix."
}'

# Third: action plan
belt app run openrouter/claude-sonnet-45 --input '{
  "prompt": "For the top 3 causes identified, provide specific A/B tests we can run to validate and fix each issue."
}'

モデル別のヒント

Claude

• 細やかな指示を得意とします
• 役割を演じる指示によく反応します
• 複雑な制約に従うのが得意です
• 明示的な出力形式を好みます

GPT-4

• コード生成に強みがあります
• 例を使うとうまく機能します
• 構造化出力が得意です
• 「段階を追って考えましょう」に反応します

FLUX

• 被写体を詳細に説明します
• スタイルの参照が効果的です
• 照明のキーワードが重要です
• ネガティブプロンプトに対応しています

Veo

• カメラの動きを表すキーワードが効果的です
• 映画的な表現がよく機能します
• 動作の説明が重要です
• 時間的な文脈を含めます

よくある間違い

間違い	問題	修正方法
曖昧すぎる	出力が予測できない	具体的な情報を追加する
長すぎる	モデルの焦点がぼやける	重要な情報を優先する
矛盾している	モデルが混乱する	矛盾を取り除く
形式の指定がない	出力に一貫性がない	形式を指定する
例がない	期待される内容が不明確	少数ショットの例を追加する

プロンプトテンプレート

コードレビュー

Review this [language] code for:
1. Bugs and logic errors
2. Security vulnerabilities
3. Performance issues
4. Code style/best practices

Code:
[code]

For each issue found, provide:
- Line number
- Issue description
- Severity (high/medium/low)
- Suggested fix

コンテンツ作成

Write a [content type] about [topic].

Audience: [target audience]
Tone: [formal/casual/professional]
Length: [word count]
Key points to cover:
1. [point 1]
2. [point 2]
3. [point 3]

Include: [specific elements]
Avoid: [things to exclude]

画像生成

[Subject with details], [setting/background], [lighting type],
[art style or photography style], [composition], [quality keywords]

関連スキル

# Video prompting guide
npx skills add inference-sh/skills@video-prompting-guide

# LLM models
npx skills add inference-sh/skills@llm-models

# Image generation
npx skills add inference-sh/skills@ai-image-generation

# Full platform skill
npx skills add inference-sh/skills@infsh-cli

すべてのアプリを表示：belt app list