social-campaign

/social-campaign — Social Content Campaign Pipeline

Orchestrator that chains 5 phases into a complete social content campaign:

STRATEGY  →  WRITE  →  REVIEW  →  DESIGN  →  SCHEDULE
 (plan)     (draft)   (audit)   (visuals)  (postiz/typefully)

Each phase is an independent block. This orchestrator is one recipe that combines them. Produces N scheduled posts with selective visuals from a single marketing goal.

When to Use

• Pre-launch content build-up (1-2 week runway)
• Ongoing content calendar creation
• Campaign bursts around announcements
• Any time you need multiple scheduled posts with visuals

Architecture

Phase 1: STRATEGY   — Plan content calendar with themes, timing, platform mix
Phase 2: WRITE      — Draft posts calibrated to real voice (not brand doc voice)
Phase 3: REVIEW     — AI slop audit + voice calibration against published posts
Phase 4: DESIGN     — Selective visual creation in Paper MCP (not every post)
Phase 5: SCHEDULE   — Route each post to postiz first, with Typefully as the X/thread specialist and fallback

Workflow

Phase 0: Load Context

1. Run mktg status --json to check brand health
2. Read brand files in parallel:
   • brand/voice-profile.md (tone, vocabulary, do's/don'ts)
   • brand/positioning.md (angles, headlines)
3. Read brand/landscape.md if it exists — check the Claims Blacklist before making ecosystem or competitive claims. If stale (>14 days) or missing, warn that market claims may be outdated.
4. Check social distribution readiness: mktg doctor --json --fields checks.postiz.configured,checks.typefully.configured
5. If Typefully is configured, load Typefully config: check default social set, list available accounts
6. If Typefully is configured, check queue schedule:
   • typefully.js queue:get --start-date <today> --end-date <+N days>
   • typefully.js queue:schedule:get (what days/times are slots)
7. Read recent published posts:
   • Prefer Typefully history when available: typefully.js drafts:list --status published --sort -created_at
   • Otherwise use the project's recent post corpus or approved drafts for voice calibration

If brand files are missing: The skill still works. Ask the user for: (1) brand voice in 2-3 adjectives, (2) target platforms, (3) content themes. Proceed with these as lightweight context. Brand files enhance, never gate.

If Typefully queue is empty (no available slots): Offer the user two options: (1) create drafts without scheduling (user schedules manually in Typefully UI), or (2) use specific datetime scheduling with --schedule "ISO_datetime". This only applies to posts routed to Typefully.

Present context summary:

Brand: loaded (voice, positioning) | OR: missing — using lightweight context
Postiz: configured for multi-provider social distribution | OR: missing — Typefully/file/browser fallbacks only
Typefully: connected for X/threads specialist routing and fallback | OR: missing
Queue: {N} empty Typefully slots over next {days} days ({schedule})
Last post: "{preview}" ({date})

Phase 1: STRATEGY

Plan the content calendar. This is strategic thinking, not template filling.

Inputs needed (use AskUserQuestion if not provided):

• What's the marketing goal? (launch, awareness, engagement, authority)
• What's the timeline? (1 week, 2 weeks, ongoing)
• Any specific themes or angles to hit?

Planning process:

1. Analyze the goal, timeline, and available queue slots
2. Plan N posts distributed across the timeline
3. For each post, decide:
   • Day and time slot (from queue schedule)
   • Platform mix (X only, LinkedIn only, or both)
   • Theme (problem awareness, value education, social proof, product tease, announcement)
   • Content type (observation, hot take, how-to, builder update, announcement)
4. Consider pacing: build gradually, don't front-load announcements
5. Consider variety: alternate themes, don't repeat the same structure

Present the calendar for approval:

| # | Date | Platform | Theme | Hook Direction |
|---|------|----------|-------|----------------|

Gate: User must approve calendar before Phase 2.

Phase 2: WRITE

Draft all posts, calibrated to the user's ACTUAL voice (not just the brand doc).

Critical: Voice Calibration

Before writing, read 5-8 recent published posts from Typefully:

typefully.js drafts:list --status published --sort -created_at
typefully.js drafts:get <id>  # for full text of top posts

Analyze the real voice patterns:

• Sentence length and structure
• Use of fragments vs complete sentences
• Technical specificity level
• How they open posts (fact-first? question? observation?)
• How they close (link? command? statement?)
• Personality markers (dry humor? directness? code inline?)

Write each post matching these patterns, NOT generic brand voice.

For each post:

1. Write the full text
2. Determine platform (X, LinkedIn, or both)
3. If cross-platform with same content, use one draft with --platform x,linkedin
4. If cross-platform with different content, plan as single draft with platform-specific text

Present all posts for approval as a numbered list with the full text.

Gate: User must approve posts before Phase 3.

Phase 3: REVIEW

Two-part quality audit. This is the most important phase for authenticity.

3a: AI Slop Audit

Audit every post for these AI tells. Read each post manually, in sequence, looking at the batch as a whole (not just individual posts). This cannot be done with regex.

Structural patterns to catch:

• Sentences that all follow the same rhythm (subject-verb-explanation, repeat)
• Every post opening with a question or every post opening with a bold claim — variety is key
• Numbered lists where every item has the same sentence structure
• "Sandwich" pattern: statement, elaboration, restatement in every post

Transition filler to remove:

• "Here's the thing:", "Let me be clear:", "The reality is:", "At the end of the day"
• "In today's [landscape/world/environment]", "When it comes to"
• "It's worth noting that", "Interestingly enough", "Not gonna lie"

Vocabulary tells to flag:

• "Leverage", "unlock", "game-changer", "deep dive", "robust", "streamline"
• "Landscape", "navigate", "elevate", "empower", "cutting-edge"
• Overuse of "incredibly", "absolutely", "fundamentally"

Batch-level monotony checks:

• Do 3+ posts start the same way? Fix it.
• Do all posts use the same CTA? Vary them.
• Is the emotional range flat? Mix insight, humor, urgency, curiosity.
• Could you shuffle the posts and not notice? Each needs a distinct angle.

If /cmo is installed, also reference ~/.claude/skills/cmo/references/ai-slop-patterns.md for the full pattern library.

3b: Voice Calibration Check

Compare each post against the published posts from Phase 2's calibration:

• Does this sound like the same person wrote it?
• Would this blend in with their real timeline?
• Is the technical specificity level right?
• Are the sentence rhythms natural?

Fix anything that fails either audit. Present the final versions.

Gate: User approves final post text.

Phase 4: DESIGN

Selective visual creation. NOT every post gets an image — text-only posts feel more authentic for personal observations and hot takes.

For the full decision framework (which posts get images), image types, Paper MCP steps, and design system specs, see references/paper-mcp-workflow.md.

Quick decision rule: Comparisons, processes, data, and announcements get images. Personal takes, vents, and builder updates stay text-only.

Present image plan for approval:

| Post # | Image? | What It Shows |
|--------|--------|---------------|

Gate: User approves image plan. User exports PNGs from Paper after design.

Phase 5: SCHEDULE

Step 0 — Pick the scheduling backend per post.

Before scheduling anything, detect which backends are available. This check runs once per Phase 5 entry and caches for the phase:

mktg doctor --json --fields checks.typefully.configured,checks.postiz.configured

Resolve the routing per post using the table below. platform is already tagged per post by Phase 2 and appears verbatim in the post file's YAML frontmatter.

Invariant: any post with type: thread stays on Typefully regardless of postiz availability. Thread UX is Typefully's canonical path; do not degrade it.

Progressive enhancement: if checks.postiz.configured === false, Phase 5 behaves exactly as the Typefully-only path below — zero behavior change for users without postiz. If both are missing, fall back to writing files into marketing/campaigns/{campaign-name}/posts/ and surface a "copy these to the platform UI" summary.

Fallback behavior under failure (during the phase):

• Postiz unreachable / 5xx: retry once with exponential backoff (1s, 2s). On second failure for a Typefully-covered platform (linkedin/threads/bluesky/mastodon), fall back to Typefully. On second failure for a postiz-only platform (reddit/instagram/tiktok/youtube/pinterest/discord/slack), fail the post with fix: "postiz unreachable — check mktg catalog info postiz --fields configured,resolved_base".
• Postiz 401: halt the entire phase. Bad key means the whole batch will fail; don't waste rate-limit quota.
• Postiz 429: parse Retry-After from the adapter's error envelope; surface to the user. Resume after the cooldown or save the batch for later.
• Typefully failure: unchanged from pre-extension behavior — fail loudly per the existing flow below.

Case A — Typefully path (X and threads, plus Typefully-covered platforms when postiz is absent)

Upload media and schedule all posts via Typefully.

For posts WITH images:

1. Find exported PNGs (ask user for location if needed)
2. Upload each: typefully.js media:upload <file_path>
3. Create or update draft with media: typefully.js drafts:create --platform x,linkedin --text "..." --media <media_id> --schedule <ISO_datetime>

For text-only posts:

1. Create draft: typefully.js drafts:create --platform x,linkedin --text "..." --schedule <ISO_datetime>

For updating existing drafts (if created earlier):

1. Upload media: typefully.js media:upload <file>
2. Update draft with text AND media: typefully.js drafts:update <draft_id> --text "..." --media <media_id> --use-default (Note: --text is required even when only adding media)

Scheduling options:

• Specific time: --schedule "2025-01-21T09:00:00Z"
• Next queue slot: --schedule next-free-slot
• Publish now: --schedule now

After scheduling all posts:

1. Verify: typefully.js drafts:list --status scheduled --sort scheduled_date
2. Confirm each draft has correct platform, text, and media
3. Present final summary table

Gate: User reviews scheduled posts in Typefully UI.

Case B — Postiz path (non-Twitter platforms when postiz is configured)

For each post whose routed backend is postiz (per Step 0 table), build a publish.json entry and invoke the adapter. You do NOT call the postiz API directly — always go through mktg publish --adapter postiz.

For text-only posts (v1 is draft-only; scheduling lands in v2):

1. Build publish.json with one item per post:

   {
     "name": "{campaign-name}",
     "items": [{
       "type": "social",
       "adapter": "postiz",
       "content": "<post text>",
       "metadata": { "providers": ["<postiz-identifier>"] }
     }]
   }
   

   providers must be exact postiz identifier strings (linkedin, linkedin-page, bluesky, reddit, etc.) — resolve friendly names via /postiz skill's Step 2b alias logic, not inline here.

2. Dry-run: mktg publish --adapter postiz --dry-run --input '<json>' --json. Verify the adapter validates the providers against connected integrations.

3. Confirm: mktg publish --adapter postiz --confirm --input '<json>' --json. The adapter writes sent-markers to .mktg/publish/{campaign-name}-postiz.json on success.

4. Update the post file frontmatter: status: scheduled, draft_id: <id>.

For posts WITH images (v2 — skip in v1): postiz v1 is draft-only; images hardcode to []. If a post has has_image: true, either (a) wait for v2, or (b) downgrade to Typefully if the platform allows (X/LinkedIn). Do NOT populate metadata.mediaPaths in v1 — the adapter ignores it.

After scheduling all posts (mixed Typefully + postiz):

1. Verify Typefully side: typefully.js drafts:list --status scheduled --sort scheduled_date.
2. Verify postiz side: the adapter's response envelope already reports per-provider status in results[]. Capture draft IDs to the post file frontmatter.
3. Present the unified final summary table.

Gate: User reviews scheduled posts in each platform's UI (Typefully for X, postiz dashboard for everything else).

Human-in-the-Loop Gates

[Phase 1] → User approves content calendar
[Phase 2] → User approves post drafts
[Phase 3] → User approves after AI slop + voice audit
[Phase 4] → User approves image plan, exports PNGs from Paper
[Phase 5] → User verifies in Typefully UI

Each gate is an AskUserQuestion. The user can go back to any phase.

Output

marketing/campaigns/{campaign-name}/
  calendar.md              # Content calendar with themes + timing
  posts/
    post-01.md             # Each post with text, platform, schedule
    post-02.md
    ...
  images/
    post-03-comparison.png # Exported from Paper
    post-06-flow.png
    ...
  campaign-log.md          # What was scheduled, draft IDs, media IDs

Post File Format

Each post-XX.md file uses this structure:

---
post_number: 1
platform: x,linkedin
type: single | thread | carousel
scheduled_date: "2026-03-25T09:00:00Z"
draft_id: ""          # Populated after Typefully draft creation
media_id: ""          # Populated after media upload
has_image: false
status: draft | approved | scheduled | published
---

[Post text goes here]

[For threads, separate tweets with ---]

Integration Requirements

If Typefully is not configured, guide setup before Phase 5. If Paper MCP is not available, skip Phase 4 (text-only campaign).

Reusable Blocks

Anti-Patterns

Principles

• Voice over brand — Match the person's real voice, not the documented brand voice
• Selective visuals — Every image must earn its place; text-only is a valid choice
• Batch coherence — Posts should vary in structure; no two adjacent posts should use the same pattern
• AI slop is non-negotiable — Manual audit every time, no shortcuts
• Skills never call skills — This SKILL.md teaches the agent the sequence; the agent loads each skill
• Filesystem is the bus — Campaign state persists as files, not in conversation
• Human gates at every phase — No runaway agent chains