variation-writer

Variation Writer Skill

Generates 3 distinct, ready-to-post LinkedIn drafts for each idea. Each uses a different hook formula AND a different post structure. Every draft includes 2 alternative hooks so the user can mix and match.

CRITICAL: Read the Best-Practice Guide First

Before writing ANY post, read references/best-practice-guide.md in this skill's directory. That guide is the operating manual. Every post must follow its principles — hooks, formatting, structure, writing tactics, and engagement mechanics. The guide contains 10 hook formulas, 6 post outlines, formatting rules, and checklists that directly shape output quality.

Triggers

• User says "yes" to an idea from idea-generator
• User runs /generate and selects ideas to develop
• User requests "regenerate" for an existing idea

Step-by-Step Instructions

Step 1: Load the Idea, Voice, and Scored Source Posts

import sqlite3, json, os, glob

# Find feeds.db
for db_path in glob.glob('/sessions/*/mnt/*/LinkedIn Feed Tracker/feeds.db'):
    break

conn = sqlite3.connect(db_path)
conn.row_factory = sqlite3.Row

# Load the idea
idea = dict(conn.execute("SELECT * FROM content_ideas WHERE id = ?", (idea_id,)).fetchone())

# Load voice config
voice = conn.execute("SELECT * FROM user_voice_config WHERE id = 1").fetchone()
if voice:
    voice = dict(voice)
    system_prompt = voice.get('system_prompt', '')
    authority_topics = json.loads(voice.get('authority_topics', '[]'))
else:
    system_prompt = ''
    authority_topics = []

# Load source posts that inspired this idea (from Plugin 1 scoring)
source_ids = json.loads(idea.get('source_post_ids', '[]'))
source_posts = []
if source_ids:
    placeholders = ','.join(['?'] * len(source_ids))
    source_posts = [dict(r) for r in conn.execute(f"""
        SELECT author_name, content, content_short, likes, comments, relevance_score, focus_areas
        FROM posts WHERE id IN ({placeholders})
    """, source_ids).fetchall()]

# Load hook performance (learning loop)
try:
    hook_perf = [dict(r) for r in conn.execute("""
        SELECT hook_type, avg_likes, avg_comments, sample_count, confidence
        FROM hook_performance
        ORDER BY confidence DESC
    """).fetchall()]
except:
    hook_perf = []

conn.close()

Step 2: Select 3 Hook Formulas + 3 Post Structures

From the best-practice guide, pick 3 DIFFERENT combinations. Never repeat a hook formula or structure across the 3 versions.

10 Hook Formulas (Section 11 of guide):

1. Curiosity Gap — hint at valuable info without revealing it
2. Quantified Promise — lead with specific number + measurable outcome
3. Contrarian / Hot Take — challenge a widely accepted belief
4. Story Opener — start with a specific, vulnerable moment
5. Problem Spotlight — name a specific pain the audience experiences
6. Bold Claim — make a confident, measurable promise
7. Data Drop — open with a surprising statistic
8. Micro-Confession — brief personal admission before pivoting to expertise
9. Framework Tease — introduce a proprietary system or mental model
10. Pattern Interrupt — break the expected pattern to force attention

6 Post Structures (Section 12 of guide):

1. Narrative Arc — Hook → Context → Conflict → Turning point → Outcome → Lesson → CTA
2. PSR (Problem-Solution-Result) — Hook → Problem → Solution → Result → Takeaway
3. Contrarian Take — Hook → Why they're wrong → What actually works → CTA
4. Tactical Listicle — Hook → Setup → 3-7 numbered points → Highlight best → CTA
5. Comparison/Contrast — Hook → X does this / Y does this → Pairs → Takeaway
6. Observation + Insight — Hook → What you're seeing → Why it's happening → What it means → CTA

Selection strategy:

• If hook_performance data exists, weight toward historically strong hooks for this user
• Always ensure diversity — 3 different hooks AND 3 different structures
• If the idea came from a focus area with high-scoring source posts, use those posts' themes as fuel (but never copy them)

Step 3: Write 3 Complete Drafts

For each version, write a complete, ready-to-copy-paste LinkedIn post plus 2 alternative hooks.

Non-negotiable quality standards (from the guide):

Hook (first 2 lines):

• Creates a curiosity gap, makes a bold claim, or names a specific pain
• Would stop the user from scrolling if they saw it in their feed
• Stays under ~150 characters to land before LinkedIn's "See more" fold

Structure:

• Line break after the hook — ALWAYS
• Max 1-2 sentences per paragraph with blank lines between every paragraph
• Total post under 3,000 characters
• Follows one of the 6 proven outlines

Writing:

• Specific over vague — concrete numbers, timeframes, names, examples
• Every claim backed by a result, data point, or named example
• Reads like the person talking, not a press release
• 6th-8th grade reading level — short sentences, common words
• Use the scored source posts' themes as fuel (reference what the network is discussing)

Engagement:

• Ends with a clear, specific CTA question (NOT "What do you think?" — something the reader can answer in one sentence)
• No engagement bait
• No external links in the post body (note to add links as first comment)

Formatting (exactly as it would appear on LinkedIn):

• Bold sparingly — 1-2 key phrases max
• Emojis: match user's emoji_usage setting (0-3 per post)
• Symbols (→ ✓ •) for visual structure, 2-3 per post max
• 3 or fewer hashtags at the bottom, or none (match user's hashtag_usage)

Voice:

• Follow the user's system_prompt voice exactly
• Match their length_pref setting
• The post must feel like something the user would actually write
• 70/30 rule: 70% personal experience/observation, 30% broader insight

Step 4: Present Each Draft with Alt Hooks

For each of the 3 versions, present:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
VERSION [N] — [HOOK FORMULA NAME] + [STRUCTURE NAME]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

PRIMARY HOOK:
[The main hook — this is the default opening]

ALT HOOK A:
[A different opening using a different formula from the 10]

ALT HOOK B:
[Another option — different angle or emotional register]

───────────────────────────────────────────

[Full post text using the primary hook, formatted exactly as it would appear on LinkedIn — short paragraphs, line breaks, symbols, ready to copy-paste]

───────────────────────────────────────────
[X] words | [Hook formula] + [Structure] | Inspired by [N] posts scoring [X]+
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

After showing all 3 versions, present the pre-post checklist from the guide (Section 13):

PRE-POST CHECKLIST:
□ Hook test — would you stop scrolling for these first 2 lines?
□ Specificity — concrete numbers, timeframes, examples?
□ One idea — single clear takeaway?
□ Format — short paragraphs, line breaks, mobile-readable?
□ Voice — read it out loud, does it sound like you?
□ CTA — clear, specific question at the end?
□ Proof — every claim backed by result or data?
□ Links — moved to first comment (not in post body)?

Then ask:

Pick a version:
- "1", "2", or "3" — Mark as your pick
- "1 with hook B" — Use version 1 but swap in alt hook B
- "blend 1+3" — Combine elements from two versions
- "tweak 2: [instructions]" — Edit a specific version
- "regenerate" — 3 new versions with different hooks/structures
- "next" — Move to the next idea

Step 5: Store Variations in Database

conn = sqlite3.connect(db_path)
cursor = conn.cursor()

cursor.execute("""CREATE TABLE IF NOT EXISTS content_variations (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    idea_id INTEGER NOT NULL,
    hook_type TEXT NOT NULL,
    structure_type TEXT,
    full_text TEXT NOT NULL,
    headline TEXT,
    alt_hooks TEXT,
    self_score REAL,
    actual_performance_json TEXT,
    created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
    posted_at DATETIME,
    FOREIGN KEY (idea_id) REFERENCES content_ideas(id)
)""")

for draft in drafts:
    cursor.execute("""
        INSERT INTO content_variations
        (idea_id, hook_type, structure_type, full_text, headline, alt_hooks)
        VALUES (?, ?, ?, ?, ?, ?)
    """, (
        idea_id,
        draft['hook_type'],
        draft['structure_type'],
        draft['full_text'],
        draft['full_text'].split('\n')[0][:80],
        json.dumps(draft['alt_hooks'])
    ))

conn.commit()
conn.close()

Step 6: Mark User's Selection

When user picks a version:

• Update content_ideas.status to 'used'
• Update content_ideas.used_at to now
• If user picked an alt hook, note which one in content_ideas.notes

Timing Recommendation

After the user picks their versions, suggest optimal posting times based on the guide (Section 7):

SUGGESTED POSTING SCHEDULE:
Based on your audience and the research:

[Day], [Time window] — [Topic/Version] — [Reason]
Example: "Wednesday, 3-5 PM — Version 1 (Story hook) — midweek afternoon peaks for B2B"

Remember:
- Space posts at least 18 hours apart
- Add a first comment within 2 minutes of posting (seed engagement, house any links)
- Stay active replying to comments for the first 60 minutes — that's the algorithm's test window

Regeneration

If user says "regenerate":

• Pick 3 NEW hook formulas (different from the originals)
• Pick 3 NEW structures (different from the originals)
• Write 3 fresh versions
• Present again with same format

Integration with linkedin-post-dev Skill

If the user has the linkedin-post-dev Cowork skill installed, the variation-writer should follow the same output conventions: 3 versions per topic, each with primary hook + 2 alt hooks, formatted for direct copy-paste. The key difference is that Plugin 2's input comes from scored feed data rather than manually provided topics.

Error States

Error	Message
Idea not found	"Idea [ID] doesn't exist. Run /generate to create ideas first."
No voice config	"Run /setup-content-engine to configure your voice before generating drafts."
Idea already used	"This idea was already used on [date]. Want to regenerate anyway?"
Best-practice guide missing	"Warning: best-practice-guide.md not found in references/. Posts will use default quality standards."