Brand Campaign Generator

Brand Campaign Generator

Visual generation (binding): all images/video route through the grounded pipeline — see .claude/rules/real-images-only.md + the grounded-visuals skill. Direct provider calls fail CI.

Purpose

Takes a Business DNA profile + a campaign goal and produces a complete set of on-brand content variations for every active platform. Outputs to the Synthex content pipeline for immediate scheduling, scoring, and publishing.

This is what Google Labs Pomelli does for visuals — but for text content with Synthex's scheduling, BYOK AI, and analytics infrastructure behind it.

Workflow

1. Load Business DNA (from brand-dna skill or ask user to describe brand)
2. Define campaign goal (product, offer, event, or theme)
3. Generate content matrix across all active platforms
4. Score each piece via content-scorer (zero AI calls)
5. Display campaign for review
6. Schedule approved content via Synthex calendar

Campaign Input Collection

Ask the user for:

Campaign Brief:
- GOAL: What are you promoting? (product, sale, event, announcement)
- DURATION: How many days? (e.g. 7, 14, 30)
- CALL TO ACTION: What should people do? (book, buy, visit, DM)
- TONE OVERRIDE: Any tone shift for this campaign? (urgent, celebratory, etc.)
- PLATFORMS: Which platforms? (or use all from Business DNA)

Content Matrix Structure

For each platform, generate:

Content Generation via Synthex Pipeline

API endpoint: POST /api/ai/generate-content

const request = {
  type: 'post',
  platform: 'linkedin', // iterate per platform
  tone: dna.toneOfVoice[0], // from Business DNA
  personaId: brandPersonaId, // if persona exists for this brand
  keywords: dna.vocabulary,
  targetAudience: dna.targetAudience,
  length: 'medium',
  includeHashtags: true,
  includeCTA: true,
};

BYOK: If user has a BYOK key, it is automatically injected via lib/ai/api-credential-injector.ts. No AI cost to Synthex.

Content Scoring

After generation, score each piece via lib/ai/content-scorer.ts. Surface scores to user in the review panel:

[LINKEDIN POST 1]
Score: 87/100  | Engagement: 90 | Platform Fit: 95 | Readability: 82
[content here]

[INSTAGRAM CAPTION 1]
Score: 79/100  | Engagement: 85 | Platform Fit: 88 | Readability: 72
[content here + hashtags]

Campaign Review Format

Present as a structured campaign block:

──────────────────────────────────────────
 CAMPAIGN: [Campaign Name]
 GOAL: [Goal]  |  DURATION: [X days]  |  POSTS: [total count]
──────────────────────────────────────────

LINKEDIN (3 posts)
  Post 1 — Score 87 ✓ [preview first 50 words...]
  Post 2 — Score 81 ✓
  Post 3 — Score 74 ⚠ [below 80 — consider revising]

INSTAGRAM (3 posts)
  ...

→ Approve all and schedule?
→ Or select specific posts to edit/replace.

Campaign visuals (REAL IMAGES ONLY)

Every image/video asset attached to a campaign post is produced ONLY via generateImage()/generateBatch() (lib/services/ai/image-generation.ts) or the generate_image / generate_video MCP studio tools — grounded-by-default on the owned reference library (public/reference-library/manifest.json; private refs via POST /api/admin/private-refs). Subjects without owned references BLOCK ("No owned references for this subject — add real photos to the reference library first") — plan real-photo ingest into the campaign timeline instead of stock or direct provider calls (banned; CI guard tests/unit/ai/no-direct-image-apis.test.ts). Carpet-cleaning campaigns auto-apply the carpet-style-v1 LoRA (trigger ccwcarpet). Use the dashboard 3-variant batch + tap-to-rank flow to select campaign imagery.

Scheduling

On approval, create scheduled posts via the scheduler API (app/api/scheduler/):

// Spread posts across campaign duration using AI-suggested timing
// lib/ai/content-repurposer.ts handles platform-specific timing rules

See social-integrations skill for platform OAuth requirements.

Reference

• Content generation: lib/ai/content-generator.ts
• Scoring: lib/ai/content-scorer.ts
• Repurposing: lib/ai/content-repurposer.ts
• BYOK injection: lib/ai/api-credential-injector.ts
• Scheduling API: app/api/scheduler/

Capability Uplift — Override Defaults

NEVER open any post with "Excited to announce", "We're thrilled to share", or "In today's fast-paced world". Never use "leverage", "game-changing", "revolutionary", "seamless", or "robust". Never produce the same tone for LinkedIn and TikTok. Never end with a three-bullet summary. Never output content that scores below 80 via the content scorer.

INSTEAD every campaign uses platform-distinct hooks:

• LinkedIn: counterintuitive insight or specific outcome in the first line
• Instagram: scene-setting or sensory detail in 125 characters
• TikTok: direct challenge or bold question in the first 3 words
• Facebook: community question or local-business relatable scenario
• X/Twitter: surprising fact or punchy opinionated take

Every piece has exactly one CTA naming the specific action AND benefit: "Book a 15-min call → get your first post live this week"

REFERENCE .claude/skills/synthex-standards/references/content-standards.md

Foundation & Gate Wiring (SYN-1049)

Adopted from the senior-skill standard so every campaign this connector produces is checked against the locked foundation before it lands.

Reads at every invocation (never cached — re-read each run):

• .claude/memory/ceo-foundation.md — universal taboos (Phase 1 + CEO standing orders), the target brand's taboos (Phase 3.X), voice tag (Q2.5.5), Phase 4 voice amendments, and the cross-client boundary (Phase 3.4) when generating CCW content.
• .claude/memory/verification-gates.md — gate state for any metric or claim a campaign references.

Output gate: every generated post routes through brand-voice-enforce before the CEO batched-review queue. A REJECT blocks that post until the quoted offending string is fixed — this is the same gate the senior-copywriter output already passes through.

Evidence standard: every quantitative claim in a campaign (scores, audience figures, performance promises) carries exactly one tag — [VERIFIED] / [INFERENCE] / [UNCONFIRMED]. Untagged = defect (.claude/rules/fabel-evidence-standard.md). Never state a projected result as fact.

Spec: see spec.md in this skill directory.