Ad Copy Generator

Write Meta ad copy that's matched to the actual image creative — not generic copy pasted across every ad. Each image gets copy that reinforces its specific message, written in the brand's voice, informed by what's already working in the account.

Read workspace/brand/ per the _vibe-system protocol if available.

Brand Memory Integration

This skill reads brand context to make every piece of copy sound like the brand, not a template.

Reads: voice-profile.md, positioning.md, audience.md (all optional)

On invocation:

1. Load voice-profile.md — Match tone, vocabulary, forbidden words. Show: "Voice loaded — [tone summary]."
2. Load positioning.md — Use angle as copy's foundation. Show: "Positioning: [angle]."
3. Load audience.md — Know who we're talking to, their awareness level, their language. Show: "Writing for [audience summary]."
4. If workspace/brand/ doesn't exist — Skip. Ask for ICP, voice notes, and forbidden words inline. This skill works standalone.

Iteration Detection

Before writing, check for existing copy:

1. Check workspace/campaigns/{name}/ads/ — If copy exists, show what's there.
2. Ask: "Revise existing, add variants, or start fresh?"
3. If nothing exists — Proceed to generation.

The Process

Step 1: Pull What's Already Working

Before writing a single word, look at the account. What copy is converting?

# Top performers by CTR — last 30 days through official Ads CLI
meta --output json --no-input ads insights get \
  --date-preset last_30d \
  --fields spend,impressions,clicks,ctr,cpc,reach \
  --sort ctr_descending \
  --limit 20

For creative details, prefer official CLI resources where supported:

meta --output json --no-input ads ad get AD_ID
meta --output json --no-input ads creative get CREATIVE_ID

If official CLI does not expose a needed creative field yet, use direct Graph API only as an explicit fallback.

Extract patterns:

• Headline length and structure (questions? numbers? commands?)
• Body copy hooks (opens with pain? data? question? story?)
• Social proof claims (X customers, Y% increase, Z revenue)
• Reading level (short punchy vs. narrative)
• CTA that converts (Learn More vs Get Started vs Book Demo)

Show the user what you found: "Your top 3 ads all open with a specific number and close with social proof. Average headline: 32 chars. I'll match that pattern."

Step 2: Load Brand Context

Read the client's CLAUDE.md, brand profile, or ask inline:

Default forbidden words (always banned unless brand explicitly uses them): revolutionary, game-changer, cutting-edge, innovative solution, transform your business, unlock your potential, unleash, harness the power, leverage (as verb), next-level, best-in-class, world-class, seamless, robust, synergy

Step 3: Analyze Each Image Creative

For every image that needs copy, use vision to identify:

1. Visual format — Notes app, receipt, tweet, chart, kitchen photo, poster, review, slack, etc.
2. On-image text — What copy is already baked into the image?
3. Angle/hook — What psychology is the image pulling?
4. Mood — Organic/found-footage, professional, urgent, casual, dark, bright
5. Funnel stage — Awareness (problem), consideration (solution), conversion (offer)

The critical rule: Copy reinforces the image, never repeats it. If the image shows "$17K/month", the body tells the story behind that number. If the image shows a stressed host, the copy describes the chaos. Image and copy are two halves of one message.

Step 4: Write Copy Using Psychology

Every ad pulls at least one psychological lever. Rotate across variants so Meta can test.

The Four Horsemen (Kallaway Framework):

Matching psychology to image format:

Step 5: Apply Copy Specs

These specs come from analyzing hundreds of winning Meta ads:

Step 6: Generate Degrees of Freedom Variants

Meta's algorithm tests combinations. Give it options:

Per creative:

• 3-5 headline variants (different angles)
• 3-5 body copy variants (different psychology)
• 1-2 description variants (supporting stat or benefit)

Rules for variant quality:

• Each variant MUST hit a different Horseman. V1=Fear, V2=Money, V3=Status, etc.
• No two headlines should use the same structure. Mix questions, statements, numbers, commands.
• No two body variants should open the same way.
• At least one variant should feel like organic content (not an ad).

Step 7: Cross-Reference with Account Winners

Before finalizing, check your copy against Step 1 findings:

• Does headline length match what's working?
• Are you using similar hook structures?
• Does reading level match?
• Are your social proof claims consistent with what the account already says?

Flag any mismatches: "Your top performers use question headlines. 2 of my 3 headlines are questions — aligned."

Funnel Stage Adjustments

Adjust body copy density:

• ToFu: 50-80 words (fast read, curiosity-driven)
• MoFu: 80-120 words (more detail, proof points)
• BoFu: 60-100 words (tight, urgent, specific offer)

Output Format

For each creative:

## [Creative Name]
*Image: [brief description of what the image shows]*
*Format: [notes app / tweet / receipt / etc.]*
*Psychology: [primary Horseman] + [secondary]*
*Funnel: [ToFu / MoFu / BoFu]*
*Matched to image: [one line on how copy reinforces the visual]*

### Headlines
1. [headline] ([XX] chars) — [Horseman]
2. [headline] ([XX] chars) — [Horseman]
3. [headline] ([XX] chars) — [Horseman]

### Body Copy

**V1 — [Horseman]: [Angle name]**
[body text]
*[word count] words*

**V2 — [Horseman]: [Angle name]**
[body text]
*[word count] words*

**V3 — [Horseman]: [Angle name]**
[body text]
*[word count] words*

### Descriptions
1. [description]
2. [description]

### asset_feed_spec (ready for API)
```json
{
  "bodies": [
    {"text": "V1 body..."},
    {"text": "V2 body..."},
    {"text": "V3 body..."}
  ],
  "titles": [
    {"text": "Headline 1"},
    {"text": "Headline 2"},
    {"text": "Headline 3"}
  ],
  "descriptions": [
    {"text": "Description 1"},
    {"text": "Description 2"}
  ]
}
`` `

Campaign File Output

After generation, save to campaign directory:

workspace/campaigns/{campaign-name}/ads/
  {creative-name}.md          <- Full copy document
  {creative-name}.json        <- asset_feed_spec JSON

Append to workspace/brand/assets.md:

| {creative-name} copy | ad-copy | YYYY-MM-DD | {campaign-name} | draft | 3 headlines, 3 bodies |

Anti-Patterns

• Reusing copy across creatives — Every image gets unique copy that matches IT
• Same psychology across variants — V1/V2/V3 must hit different Horsemen
• Starting with brand name — Nobody cares about your brand in ToFu
• Generic SaaS language — See forbidden words list
• Emoji bullet lists — Looks like every other ad on Meta
• Ignoring account data — Always check what's working first
• Copy that repeats the image — Image says the number, copy tells the story
• Long paragraphs — 3+ sentences = scroll past on mobile
• Same headline structure — Mix questions, stats, commands, statements
• Hard sell CTAs in ToFu — "Sign up now" in awareness = low CTR

Quick Mode

If the user provides images and says "write copy for these" without specifying a full workflow:

1. Skip account data pull (or use cached if available)
2. Load brand context if it exists
3. Analyze images → write matched copy → output in standard format
4. Still suggest checking against account data as a next step

Integration

This skill connects to the Meta Ads Copilot ecosystem:

• Upstream: ad-creative-monitor detects fatigue → triggers copy refresh
• Downstream: Copy output feeds into Meta Graph API ad creation/update
• Learnings loop: Track which copy patterns win, append to workspace/brand/learnings.md

Saved: workspace/campaigns/{name}/ads/ + workspace/brand/assets.md

Next up: Push copy to Meta via Graph API (ad-upload skill) or review in Ads Manager manually.