markus the open-source AI marketer

GMAC Member

GMA Community

"i gave my openclaw agent markus (the marketer) control of 8 social media accounts across 4 different use cases and let it run for 18 days. every night it pulled its own analytics, figured out what worked, improved its content strategy, and generated the next day's posts. 145 posts total, zero manual scheduling. I'm open-sourcing the entire pipeline below. for context: i've built 20+ MVPs for early-stage startups since 2021 (https://www.jclvsh.art/jclvsh). this year i'm building one product every two months - one month dev, one month gtm. the first one is dropspace (https://www.dropspace.dev) - multi-platform content distribution for AI agents via REST API and MCP. between the products, a music community i run in nyc, and my own artist projects, that's 8 social accounts across 4 platforms. maintaining all of that by hand would eat 20+ hours a week, so i built an openclaw automation pipeline on top of the dropspace API. here's how the whole thing works. one config file per use case every use case gets an `app.json` - identity, platforms, posting times, API keys, integrations, notification routing. the pipeline reads it and adapts. spinning up a new use case is one command: node init-app.js --app myapp --platforms tiktok,twitter,linkedin the repo has three starter templates ready to fork: dropspace-content-engine - the full autonomous loop. your agent scans X for trending hooks, analyzes 14 days of post performance, generates platform-specific content (slideshows, tweets, linkedin posts, reddit threads), and schedules via the API. every night the cycle repeats with fresh data from the previous day's posts. dropspace-dj-clipper - for anyone who records long audio or video. the pipeline analyzes a recording, identifies transitions using audio peak detection and whisper-based timestamping, cuts 30-second vertical clips, generates captions with artist credits, and schedules them across tiktok and instagram. dropspace-photo-slideshows - for communities and venues that run events. event photos go into google drive, the pipeline assembles tiktok slideshows with face-aware text overlays, generates captions in a configurable voice, and schedules one post per day. the nightly pipeline four crons run sequentially every night: midnight - pulls analytics for all recent posts across all use cases via the dropspace batch analytics API and cleans up image assets older than 7 days. x-research (12:30 AM ET) - uses bird CLI to scan X for trending hooks in your niche across 13 search queries plus 5 competitor handles, then distributes signals to a shared file that self-improve reads. self-improve (1:00 AM ET) - the core engine. it pulls 14 days of engagement data per platform, computes which hooks drove clicks and which formats got views and which posting times performed, then cross-references all of that with x-research signals. an LLM ingests the data plus persistent strategy notes from previous runs and generates complete post blueprints. the queue caps at 14 posts per platform and auto-refills when it drops below 7, with a fact-checking step that verifies any claims about people, products, or events before saving. schedule-day (2:00 AM ET) - reads posting times from `app.json`, applies ±15 min random jitter so posts don't land at identical times, and creates scheduled launches via the API. visual platforms get images generated through fal.ai with text overlays rendered by node-canvas, text platforms get content formatted per platform rules from a shared config registry (`platforms.js`). march 2–20: 4 use cases i ran this pipeline across 4 projects from march 2 through march 20. the aggregate: 145 posts, 28,092 impressions, 318 reactions, 42 comments. dropspace (https://www.dropspace.dev) (autonomous) - 113 posts, 13,488 impressions, 0.4-0.5% engagement. the agent generates everything with zero human review, and all brand accounts started from zero followers. jclvsh (https://www.jclvsh.art/jclvsh) (slack approval) - 16 posts, 10,957 impressions. linkedin averaged 1,057 impressions per post with 11.1 reactions. the agent drafts from my content ideas + project data (github commits, posthog traffic, sentry errors, stripe revenue) and i edit and approve in slack - about 30 seconds per post. ynho (https://open.spotify.com/artist/...) (hand-picked clips) - 10 posts, 2,085 impressions, 5.0-7.6% engagement. the agent analyzes 90-minute DJ sets via audio peak detection, cuts clips, and identifies songs. i pick which clips go out. studio jams (https://studiojams.vercel.app) (google drive photos) - 6 posts, 1,562 impressions, 1.5% engagement. event photos go into a drive folder and the pipeline assembles slideshows with face-aware text overlays and schedules them. where the gap comes from the engagement range was 0.4% to 7.6% from the same pipeline - same API calls, same scheduling, same agent. two variables moved together: how much human curation sat between the agent and the publish button, and how many existing followers each account had. ynho had ~50-90 followers and hand-picked clips (7.6%). dropspace had zero followers and zero human review (0.4%). i can't cleanly isolate which variable mattered more - both contribute. the pattern that was consistent across every use case: the automation handles scheduling, formatting, posting, and analytics while you decide what's worth amplifying. the repo 91 files with setup guides for OpenClaw, Claude Code, Cursor, and Manus. engines for visual posts (tiktok/instagram slideshows), text posts (twitter/linkedin/reddit/facebook), and video posts (ugc-reaction clips). a clipper module for DJ sets and long recordings. a slideshow builder for photo content. shared core libraries for API calls, path resolution, platform config, overlay rendering, and media generation. the full pipeline, the templates, and the case study with all the data: case study: https://www.dropspace.dev/case-s... community templates: https://www.dropspace.dev/community openclaw skill: https://clawhub.ai/jclvsh/dropspace open source: https://github.com/joshchoi4881/..."