GitHub Actions Automation Pipeline: From Blog Posts to AI Video Generation
Running a solo dev project means you can't afford to do the same thing twice. I've automated blog publishing, video generation, competitor monitoring, and infrastructure health checks via GitHub Actions. Here's the architecture.
Full Workflow Map
Daily 06:00 JST
├── daily-report.yml → KPI fetch → Slack notification
├── cs-check.yml → pending tickets → AI reply
└── ai-university-update.yml → RSS feeds → DB update
Weekly Sunday JST
├── evaluate-predictions.yml → horse racing accuracy evaluation
└── weekly-sns-draft.yml → X post draft generation
Manual / PR-triggered
├── blog-publish.yml → dev.to + Qiita post
├── deploy-prod.yml → Firebase Hosting deploy
└── video-pipeline.yml → NotebookLM → ElevenLabs → video
blog-publish.yml: The Orphan Branch Pattern
on:
workflow_dispatch:
inputs:
draft_path:
description: 'JA draft path'
draft_path_en:
description: 'EN draft path'
platforms:
default: 'devto'
dry_run:
default: 'false'
The key design decision is the orphan branch pattern:
- name: Update published:true
run: |
sed -i 's/^published: false/published: true/' "${{ inputs.draft_path }}"
git commit -m "published: ${{ inputs.draft_path }}"
git push origin HEAD:blog-publish/${{ github.run_id }}-$(date +%Y%m%d-%H%M%S)
The published: true update goes to a dedicated branch, not main. Claude Code merges it after verifying the post URL. This pattern avoids branch conflicts with parallel instances.
video-pipeline.yml: Fully Automated AI Video
steps:
- name: Generate script via NotebookLM
run: notebooklm ask "$TOPIC" > script.md
- name: Generate audio via ElevenLabs
run: |
curl -X POST "https://api.elevenlabs.io/v1/text-to-speech/$VOICE_ID" \
-H "xi-api-key: $ELEVENLABS_KEY" \
-d "{\"text\": \"$(cat script.md)\"}" \
> audio.mp3
- name: Render video via Remotion
run: npx remotion render VideoTemplate --props='{"audioFile":"audio.mp3"}'
- name: Upload to Supabase storage
run: supabase storage upload videos/$(date +%Y%m%d).mp4 output/video.mp4
NotebookLM generates the script → ElevenLabs converts to audio → Remotion renders the video. Zero manual steps.
cs-check.yml: AI Customer Support
on:
schedule:
- cron: '0 */6 * * *' # every 6 hours
steps:
- name: Get pending tickets
run: |
TICKETS=$(curl -s "$SUPABASE_URL/functions/v1/get-support-tickets" \
-H "Authorization: Bearer $SERVICE_KEY")
- name: AI reply via Claude Haiku
run: |
echo "$TICKETS" | claude --model claude-haiku-4-5 \
"Reply to these support tickets. Be helpful and specific."
Every 6 hours: check tickets → Haiku drafts reply → EF posts it. Using Haiku keeps the cost minimal.
Three Design Principles
1. Claude-independent design for cron tasks.
Scheduled workflows don't depend on Claude API availability. API outages don't interrupt operations. Haiku is used for reply drafts, but design/judgment tasks are separated.
2. dry_run on every dispatch workflow.
Every manually dispatched workflow has a dry_run input. New workflows are always verified dry before live.
3. Right concurrency settings per workflow.
deploy-prod uses cancel-in-progress: false (queue, don't drop). blog-publish allows cancellation (duplicate-post prevention is handled by the pre-check step, not concurrency config).
The Result
After full automation, my time on routine operations dropped to near zero. Blog posts, CS responses, competitor monitoring, video generation — all handled by cron. What remains: deciding what to build next. That's the CEO-only model applied to solo dev.