insightpulseai/agents/skills/social/caption_writing

Skill: Caption Writing

Skill ID: social.caption_writing Agent: SocialMediaManager (agent_008) Version: 1.0.0

What It Is

Writes platform-native captions, post copy, and thread content. Adapts tone, length, format, and CTA style to each platform's conventions and audience expectations.

When to Use

• Drafting individual posts for any social platform
• Writing threads or carousel copy
• Repurposing long-form content into social snippets
• Generating caption variations for A/B testing

Core Pattern

Input: Topic + key messages + platform + brand voice + campaign objective
Process:
  1. Select tone modifier based on platform and content type
  2. Draft copy within platform character limits
  3. Craft CTA aligned to campaign objective and funnel stage
  4. Generate hashtag set with relevance scoring
  5. Run brand_voice_enforcement check
  6. Run platform_fit check
Output: Platform-ready caption with hashtags, CTA, and quality scores

Platform Specifications

| Platform | Max Chars | Hashtag Limit | CTA Style | |----------|-----------|---------------|-----------| | LinkedIn | 3,000 | 3-5 | Professional, link-embedded | | X/Twitter | 280 | 1-2 | Direct, thread-continued | | Instagram | 2,200 | 10-15 (of 30 max) | Link-in-bio reference | | Facebook | 63,206 | 2-3 | Inline link, comment CTA | | TikTok | 4,000 | 3-5 | Overlay text, comment pin |

Copy Structure Templates

Hook → Value → CTA (Universal)

[Hook: question, bold statement, or stat]

[Value: insight, story, or framework]

[CTA: specific action request]

[Hashtags]

Thread Format (X/Twitter, LinkedIn)

1/ [Hook — the reason to keep reading]
2/ [Context — set the scene]
3-N/ [Value points — one per post]
N+1/ [Summary + CTA]

Visual-First (Instagram, TikTok)

[Caption supports the visual — don't repeat what's obvious in the image/video]
[Add context the visual can't convey]
[CTA]
[Hashtags below fold]

A/B Variant Generation

When variant_config is provided, generate multiple variants per post:

Input: Base message + variant_config
Process:
  1. Generate base version using top-performing past posts as few-shot examples
  2. For each vary_dimension:
     a. Create variant keeping all other dimensions constant
     b. Only vary ONE dimension per variant (for clean A/B signal)
  3. Score each variant with content_quality_judge
  4. Rank by predicted engagement (informed by PostPerformanceRecord patterns)
Output: VariantSet with recommended_variant_id

Variant Dimensions

| Dimension | What Changes | Example | |-----------|-------------|---------| | hook | Opening line style | Question vs. statistic vs. bold claim | | tone | Voice intensity | Authoritative vs. conversational vs. urgent | | cta | Action request | "Download now" vs. "See the data" vs. "Join the conversation" | | length | Copy length | Concise (50 chars) vs. standard (150) vs. long-form (500+) | | format | Post structure | Single post vs. thread vs. list format |

Few-Shot Example Selection

When historical performance data is available:

1. Filter PostPerformanceRecord[] by matching platform + content_pillar
2. Sort by engagement_rate_7d descending
3. Select top 3-5 as few-shot examples for generation context
4. Include the full copy + metadata so the model learns patterns, not just topics

Quality Checks

• Character count within platform limit
• CTA present and aligned to objective
• Brand voice alignment ≥ 0.85
• No do-not-say list violations
• Hashtags relevant and not banned
• No broken or placeholder links
• If variants generated: each varies exactly one dimension from base