The first time I shipped an open source tool, I wrote a good README, posted to Hacker News, got three upvotes, and concluded the project wasn't useful. It was useful. I just never showed up again.
Distribution for OSS isn't a one-day task. It's submitting to directories, writing posts across four platforms, rotating angles so you're not repeating yourself, and doing all of that every week without burning out on it. Most people don't have the time or the inclination, so the project dies.
I built marketing-pipeline to run that recurring work without daily intervention. One command onboards a project:
marketing onboard --name my-tool --repo owner/repo --kind mcp-server
That fetches the README, sends it to Claude, and writes problem statements, factual claims, and rotation angles to projects.yml. From there, a GitHub Actions cron at 14:00 UTC on weekdays picks the least-recently-used angle per project, drafts posts for Bluesky, Mastodon, X, Dev.to, and Hashnode, enforces per-channel length limits (Bluesky 300 chars, X 280, Dev.to 150-400 words), and publishes.
The part I want to highlight here is the anti-slop gate in pipeline/antislop.py.
LLM-drafted marketing copy has a fingerprint: "excited to share," "game-changer," "unlock," "empower," "AI-powered," rhetorical questions, emoji, hashtags, exclamation points. These tokens don't just read badly — they signal that nobody with actual opinions wrote this. The antislop gate hard-rejects any draft containing them before it ever reaches a publish step. This is a regex blacklist, not a prompt instruction. Prompting Claude to "sound natural" doesn't hold over time; a regex gate does.
Directory submission is also automated by project kind: mcp-server routes to MCP Registry, Smithery, Glama, and PulseMCP; claude-skill to awesome-claude-code (payload generated, but their rules require a human to submit via GitHub issue — that one step can't be scripted out); browser-extension to Chrome Web Store, Firefox AMO, and Edge Add-ons.
Top comments (0)