Marketing content generation

Two flows in one recipe.

Reactive: a marketer drops assets (images, brand-voice doc, product screenshots, transcripts) into chat with a one-line brief. The agent pulls the right skills from skills.sh, drafts a script in the brand voice, generates voiceover with ElevenLabs, renders a Remotion video, and returns a ready-to-post bundle.

Proactive: every morning at 06:00, a scheduled task walks the knowledge graph for fresh signal — new GitHub releases, customer wins from HubSpot, viral support threads in Slack — and pitches 3 concrete content ideas in #marketing-ideas with a "build it" button next to each.

15MarketingMarketing content generation

Context Graph tribal knowledge

brand voice · prior posts · rejected ideas · voice IDs

Sources

Figma

Remotion

ElevenLabs

brief + assets

Agentcy

Content agent + skills

video bundle

Output

Video

+ scripts

At a glance

Inputs: dropped assets + brief (chat) or the Knowledge Graph + skills.sh registry (cron)
Trigger: chat message with attachments or cron daily 06:00 UTC
Output: MP4 video + script.md + social copy + ready-to-post snippet pack — all to Drive/Slack
Gates: drafts are auto; publishing to public surfaces requires approval

Stack

Layer	What we use
Asset intake	Chat message attachments, Google Workspace Drive folder, Figma for design tokens
Reasoning	Agent loop with skills — fetched dynamically from skills.sh registry
Voiceover	ElevenLabs — 4 tools (text-to-speech, list voices, voice settings, usage)
Video render	Remotion inside a CIAB sandbox — bundled `slideshow` / `intro` / `graph-recap` templates or a Git project
B-roll	Runway AI — text→video, image→video, text→image
Distribution	Slack, Drive, optional CMS write-back
Scheduling	Cron task for the proactive flow
Memory	The agent remembers your brand voice decisions, voice IDs, and previously-rejected ideas so it doesn't suggest them again

Skills powering the recipe

The agent doesn't ship with marketing knowledge baked in — it pulls skills from skills.sh on demand using search_skills_registry + fetch_registry_skill, or installs them up-front via the Skills page. Real skills with real install counts you can use today:

Skill	Source	Installs	Use for
`copywriting`	`coreyhaines31/marketingskills`	79k+	Headlines, body copy, CTAs
`copy-editing`	`coreyhaines31/marketingskills`	47k+	Tightening drafts
`social-content`	`coreyhaines31/marketingskills`	54k+	Platform-specific framing
`seo-audit`	`coreyhaines31/marketingskills`	88k+	Keyword + structure checks
`marketing-psychology`	`coreyhaines31/marketingskills`	58k+	Hooks, persuasion patterns
`script-writer`	`ailabs-393/ai-labs-claude-skills`	—	Video / shorts scripts
`brand-voice` / `brand-voice-enforcement`	`anthropics/knowledge-work-plugins`	—	Stays-on-tone editor pass
`youtube-thumbnail-design`	`infsh-skills/skills`	3.9k+	Thumbnail composition
`programmatic-seo`	`coreyhaines31/marketingskills`	52k+	Bulk landing-page templates

The recipe enables a small core set in the system prompt and lets the agent fetch others on demand within the prompt budget (default 16 000 chars).

What you'll build — Reactive flow

A marketer drops a folder of assets and a brief into the agent. The agent:

Inventory — lists what was uploaded (logo, 4 product screenshots, a customer-call transcript, a brand-voice PDF). Acknowledges each.
Brief understanding — confirms the format ("30s vertical Instagram Reel"), audience ("CTOs at fintechs"), and the one thing it should land.
Skill fetch — search_skills_registry({ query: "vertical reel script CTO"}) → fetch_registry_skill(script-writer) + brand-voice-enforcement. Agent confirms char budget.
Script draft — written in the brand voice, with timecodes for visuals + voiceover.
Voice — elevenlabs_list_voices → user picks (or recipe defaults to a remembered voice). elevenlabs_text_to_speech produces the VO track.
B-roll — for any timecode without a provided asset, agent calls runway_text_to_image or runway_text_to_video from prompts grounded in the script.
Render — Remotion source is the slideshow template (or a Git project the team maintains), filled with assets + VO + brand tokens from Figma. CIAB renders MP4.
Bundle — drops the MP4, the script.md, a social copy pack (LinkedIn / Twitter / IG caption variants from social-content skill), and a thumbnail (youtube-thumbnail-design).

shaked  [drops 5 files] make a 30s vertical reel for CTOs about our new
        observability feature. Brand voice doc is in there. Audience is
        SREs at series-B fintechs.

agentcy ✅ Inventory
        - logo.svg
        - 3 product screenshots (panel-rendering)
        - customer-call.txt (Sumant from Lattix, 14 min, June 8)
        - brand-voice-v3.pdf

        ⚙ Fetching skills
        ✓ script-writer (1,847 chars)
        ✓ brand-voice-enforcement (2,103 chars)
        ✓ social-content (1,420 chars)
        Budget: 5,370 / 16,000 chars

        📝 Script draft (v1) — 28s, 6 scenes
        [00:00] Hook: "your dashboards are lying to you" (logo wipe)
        [00:03] Pain: 3 quick clips of frustrated SRE staring at tabs
        ...

        Continue? Y to render, or comment to revise.

shaked  good but tighten 00:18-00:24, and use Sarah's voice

agentcy ⚙ Tightening, switching to voice "Sarah" (id: pNInz...)
        ✓ Revised script
        ✓ ElevenLabs render: 28.4s VO (eleven_multilingual_v2)
        ✓ Runway b-roll: 2 clips for SRE-pain scenes
        ⚙ Remotion render (CIAB sandbox)
          → 1080x1920 H264 30fps
          → 28.4s · 12.4 MB

        ✅ Done
        🎬 reel-observability-2026-04-26.mp4
        📄 script.md
        📱 social-pack.md (LI, X, IG)
        🖼 thumbnail.png

        Publish? (sends to #marketing-publish for approval)

What you'll build — Proactive flow

A scheduled task runs each morning. It walks the Knowledge Graph for activity that's worth talking about and pitches content ideas:

yaml

name: marketing-ideas-daily
schedule: "0 6 * * *"
realm: marketing
agent: content-strategist
enabled_skills:
  - marketing-psychology
  - social-content
  - copywriting
prompt: |
  Walk the knowledge graph for the last 24 hours. Surface candidates worth turning into content:
    1. graph.search({ realm: ["product","gtm"], since: "24h" })
       — new GitHub releases, merged PRs touching customer-facing features
    2. hubspot.search_deals({ stage: "closed-won", since: "7d" })
       — fresh customer wins (anonymize unless tagged "case-study-ok")
    3. slack.list_messages({ channels: ["#wins","#cs-leads"], since: "24h" })
       — viral or notable threads
    4. grafana_query — usage spikes worth a "scaling lessons" post

  For each candidate, draft:
    - hook (≤ 12 words)
    - 1-line angle
    - format suggestion (reel, blog, thread)
    - which skills you'd pull and why

  Post 3 picks (NOT all candidates — be opinionated) to #marketing-ideas
  with a 🚀 reaction button. Reaction → builds the content end-to-end via
  the reactive flow above.

  Memory:
    - Read `rejected_ideas` and `recent_topics` to avoid repeats and over-coverage.
    - On 🚀 react, write the resulting brief back as `published_idea`.

What lands in #marketing-ideas at 06:05:

agentcy 🌅 3 ideas for today (Mon Apr 26)

        1️⃣ "we shipped panel rendering. it's faster than your dashboard."
            angle: real numbers from prod metrics — concrete, no hype
            format: 30s vertical reel + LinkedIn thread
            skills: copywriting, marketing-psychology, video-script-writer
            evidence: PR #2103 (merged Sun), Grafana p99 280ms→190ms
            🚀 build it

        2️⃣ "Lattix went from 4 tools to 1 in 3 weeks"
            angle: customer-side win, anonymized unless permission
            format: 90s case-study video + Twitter thread
            skills: brand-voice, social-content
            evidence: closed-won deal · Sumant call transcript
            🚀 build it

        3️⃣ "the observability tax: how we cut our Datadog bill 40%"
            angle: contrarian engineering POV, evergreen
            format: long blog post + LinkedIn carousel
            skills: copywriting, programmatic-seo, seo-audit
            evidence: Internal cost-watch dashboard, FinOps recipe data
            🚀 build it

A 🚀 reaction triggers the reactive flow with the chosen idea as the brief — the agent goes from idea to ready-to-publish bundle without anyone touching keys.

Prerequisites

1 · ElevenLabs

bash

curl -X POST $API/api/v1/sources \
  -H "authorization: Bearer $TOKEN" -H 'content-type: application/json' \
  -d '{
    "name": "elevenlabs-prod",
    "connector": "elevenlabs",
    "realm": "marketing",
    "config": { "api_key": "sk_..." }
  }'

Test with elevenlabs_list_voices to make sure your voices show up.

2 · Remotion + CIAB

The Remotion connector renders inside a CIAB sandbox. Make sure CIAB is enabled (AGENTCY_FEATURES_CIAB=true) and the EC2 runtime or local Docker runtime is configured.

bash

curl -X POST $API/api/v1/sources \
  -H "authorization: Bearer $TOKEN" -H 'content-type: application/json' \
  -d '{
    "name": "remotion-marketing",
    "connector": "remotion",
    "realm": "marketing",
    "config": {
      "project": { "kind": "git", "repo": "acme/marketing-remotion", "ref": "main" },
      "render_defaults": { "fps": 30, "codec": "h264", "resolution": "1080p" }
    }
  }'

The git project type points at your team's Remotion compositions. The bundled slideshow / intro / graph-recap templates work as starters but a custom Git project quickly pays for itself — that's where your brand-tuned compositions live.

3 · Runway (optional, for B-roll)

bash

curl -X POST $API/api/v1/sources \
  -H "authorization: Bearer $TOKEN" -H 'content-type: application/json' \
  -d '{
    "name": "runway-marketing",
    "connector": "runway",
    "realm": "marketing",
    "config": {
      "api_key": "rw_...",
      "default_model": "gen3",
      "default_duration": 4,
      "default_ratio": "9:16"
    }
  }'

4 · Figma (optional, for brand tokens)

If your brand uses Figma variables for colors / typography, the Figma connector exposes them so Remotion compositions stay on-brand without manual color hex juggling.

5 · Skills configured

In Settings → Skills:

skills_dynamic_search_enabled: true — lets the agent fetch from skills.sh on demand
skills_max_prompt_chars: 16000 (default is fine for marketing)
skill_enablement_policy: prompt-user — so dynamic skills require approval the first time

Pre-install the core pack via the Skills → Browse tab:

coreyhaines31/marketingskills — the workhorse pack
anthropics/knowledge-work-plugins/brand-voice (or brand-voice-enforcement)
ailabs-393/ai-labs-claude-skills/script-writer

Worked example — the policy

rego

# policies/content-generation.rego
package agentcy

# Drafting and rendering: free.
default allow := false

allow if {
  input.tool in {
    "elevenlabs_list_voices", "elevenlabs_get_voice_settings",
    "elevenlabs_text_to_speech", "elevenlabs_get_usage",
    "remotion.render", "remotion.list_compositions",
    "runway_text_to_video", "runway_image_to_video", "runway_text_to_image",
    "figma.get_file", "figma.get_components", "figma.get_styles",
    "google_drive.list", "google_drive.read_file", "google_drive.upload",
    "graph.search", "graph.related",
    "scenario.create",
    "search_skills", "enable_skill", "disable_skill", "list_enabled_skills",
    "search_skills_registry", "fetch_registry_skill",
  }
}

# Publishing to public surfaces requires explicit approval.
allow if {
  input.tool in {
    "slack.post_message",
    "twitter.create_post",
    "linkedin.create_post",
    "youtube.upload_video",
  }
  input.approval.granted
}

# Rate-limit ElevenLabs to avoid runaway TTS costs.
deny[msg] if {
  input.tool == "elevenlabs_text_to_speech"
  count(input.recent_calls_60s) > 10
  msg := "rate limit: 10 ElevenLabs renders/min — batch your script first"
}

# Don't publish content referencing a closed-lost deal as a customer win.
deny[msg] if {
  input.tool in {"twitter.create_post", "linkedin.create_post"}
  some claim
  claim := input.args.claims[_]
  claim.kind == "customer_reference"
  claim.deal_stage == "closed-lost"
  msg := "cannot reference a closed-lost deal as a customer win"
}

Memory — the agent gets sharper over time

rust

// Agent decides these are durable
remember("brand_voice_pref",   "voiceover: Sarah (pNInz...) for product, Dan for technical")
remember("rejected_idea",      "the 'we cut Datadog 40%' angle — legal flagged comparative claims")
remember("recent_topic",       "panel rendering shipped Apr 24 — covered in 2 reels, 1 blog. cooldown until May.")
remember("published_idea",     "Lattix case study — IG reel + LI carousel, posted Apr 26")

Recall (stage 2a of the Agent Loop) means tomorrow's proactive run won't pitch the panel-rendering angle again, will offer Sarah's voice for the product reel, and won't re-suggest the Datadog comparison.

Distribution

Three modes, opt in:

Drive bundle — MP4 + script.md + social-pack.md uploaded to a Drive folder per campaign. Manual posting from there.
Slack approval channel — bundle dropped in #marketing-publish; reactions trigger publication: 📱 → IG, 🐦 → X, 💼 → LinkedIn.
CMS / scheduler write-back — push to Buffer, Hootsuite, or your own CMS via the OpenAPI connector.

What good looks like

A reactive run completes in 2-4 minutes for a 30s vertical reel: ~10s for skill fetch, ~30s for script draft, ~20s for ElevenLabs TTS, ~60s for Runway b-roll, ~30s for Remotion render. The marketer's only job: read the script, react to publish.

A proactive day looks like 3 thoughtful suggestions at 06:05 grounded in real activity, with a 1-click 🚀 path to a finished bundle. Over time, memory + audit log shows which suggestions actually shipped — feeds back into the strategist's prompt.

Variations

Long-form blog mode — swap Remotion render for a Markdown bundle + Hero image (Runway). Same skills, different output. Good for evergreen SEO via programmatic-seo.
Per-channel sub-agents — dispatch sub-agents (one per platform: IG, X, LinkedIn, YouTube). Each has its own skill set and tone. Parent agent assembles the campaign brief.
Voice-first review — bind to the voice channel. Marketer drives via voice while reviewing generated drafts on a tablet.
Internal comms variant — same recipe, different realm. Pitch internal-comms videos (org change announcements, security awareness) to People & Comms instead of GTM.
Customer-success replay — auto-generate a "you shipped this in 30 days" video for new customers from their actual usage data — sales enablement gold.

Troubleshooting

Skills budget exhausted. The agent will refuse to enable an additional skill — disable an unneeded one (disable_skill) or bump skills_max_prompt_chars in Settings.
ElevenLabs voice not found. elevenlabs_list_voices returned only premade voices. Custom voices live under your account — the API key needs the right scope, and the voice has to be cloned into your library first.
Remotion render fails with "module not found". The Git project at repo:ref is missing a dependency. Check the CIAB sandbox build logs — npm install runs before render, log lines are tagged remotion.install.
B-roll looks generic. Runway's gen3 model needs specific prompts. The agent's default is conservative — add a b-roll-prompting skill to the active set or pass prompt_style: "specific" in the args.
Proactive task fires but says "no candidates". The graph queries returned empty — likely no fresh activity in the realms it's looking at, or the cron's since window is too narrow. Widen to since: "72h" for a sanity check.
The agent keeps re-pitching the same idea. Memory not writing. Check agentcy-memory logs — remember(rejected_idea, ...) should appear after each rejection. If it isn't, the agent isn't being told to write rejections; add it to the prompt.

What's deferred

❌ Auto-publishing without human approval — every public post requires explicit approval in v1
❌ Native Buffer/Hootsuite connectors — use OpenAPI connector for now
❌ Live A/B testing of generated content — analytics flow back is on the roadmap
❌ Cursor/Codex sub-agents inside CIAB — Claude Code only today

Concept: Skills — how the prompt budget works, skills.sh registry
Connector: Remotion — bundled templates, Git projects
Connector: ElevenLabs — voice cloning, voice settings
Connector: Runway — text/image-to-video
Concept: Tasks & Workers — the proactive cron
Recipe: Executive Reporting — sibling proactive recipe over the same graph

Marketing content generation ​

At a glance ​

Stack ​

Skills powering the recipe ​

What you'll build — Reactive flow ​

What you'll build — Proactive flow ​

Prerequisites ​

1 · ElevenLabs ​

2 · Remotion + CIAB ​

3 · Runway (optional, for B-roll) ​

4 · Figma (optional, for brand tokens) ​

5 · Skills configured ​

Worked example — the policy ​

Memory — the agent gets sharper over time ​

Distribution ​

What good looks like ​

Variations ​

Troubleshooting ​

What's deferred ​

Next ​

Marketing content generation

At a glance

Stack

Skills powering the recipe

What you'll build — Reactive flow

What you'll build — Proactive flow

Prerequisites

1 · ElevenLabs

2 · Remotion + CIAB

3 · Runway (optional, for B-roll)

4 · Figma (optional, for brand tokens)

5 · Skills configured

Worked example — the policy

Memory — the agent gets sharper over time

Distribution

What good looks like

Variations

Troubleshooting

What's deferred

Next