Skip to content

README.md

Latest commit

 

History

History
85 lines (62 loc) · 3.65 KB

README.md

File metadata and controls

85 lines (62 loc) · 3.65 KB

Content Engine

Autonomous content pipeline that captures what I'm working on, generates posts across platforms, and learns from what actually performs.

I built this because I was spending 20 minutes per LinkedIn post doing the same thing every time — pulling context from session notes, structuring it against my voice, posting, then never looking at what worked. That's a system, not a creative act. So I automated the system and kept the creative part.

What it does

  • Captures daily context from my Claude Code session history and project notes — what I shipped, what I learned, what failed
  • Generates content using semantic blueprints that encode my voice, brand constraints, and platform rules
  • Posts to LinkedIn via OAuth with draft review, scheduling, and approval gates — nothing goes out without my sign-off
  • Tracks engagement and feeds performance data back into the pipeline so the system learns what lands

Current status: Phase 3 complete (26/26 stories). Semantic blueprints, multi-agent validation, and content generation all working. Feedback loops (Phase 6) are next.

Architecture

Context Capture  →  Blueprint Engine  →  Multi-Agent Pipeline  →  Platform Posting
(sessions, notes)   (STF, MRS, SLA, PIF)  (generate → validate → refine)  (LinkedIn, drafts, scheduling)
                                                    ↑                              │
                                                    └──── engagement feedback ─────┘

The pipeline uses a Generator → Validator → Refiner pattern. Llama drafts, Claude validates against brand constraints, refinement happens in-loop. Blueprints aren't prompts — they're structured decision frameworks that encode what makes a post mine vs. generic AI slop.

Tech stack

  • Python 3.11+ with uv for package management
  • AI: Ollama locally (llama3:8b, free) → AWS Bedrock in production (Claude Haiku + Llama 3.3 70B, ~$0.004/post)
  • Database: SQLite dev → PostgreSQL production, Alembic migrations
  • Testing: 403 tests passing, ruff compliant, mypy typed
  • Deployment: Self-hosted, ./scripts/deploy.sh

Quick start

git clone <repo-url> && cd ContentEngine
uv sync
cp .env.example .env   # add LinkedIn credentials
uv run alembic upgrade head
uv run content-engine capture-context
uv run content-engine generate --pillar what_building --framework STF

Key commands

# Context capture
uv run content-engine capture-context              # today's context from sessions + notes
uv run content-engine capture-context --date 2026-01-12

# Content generation
uv run content-engine blueprints list              # available frameworks
uv run content-engine generate --pillar what_building --framework STF
uv run content-engine sunday-power-hour            # batch workflow for the week

# Post management
uv run content-engine draft "Your content here"
uv run content-engine approve 1                    # review then post
uv run content-engine approve 1 --dry-run          # test without posting

# Analytics
uv run content-engine collect-analytics
python scripts/analytics_dashboard.py

Roadmap

  • Phase 1: LinkedIn OAuth + posting infrastructure
  • Phase 1.5: Database, CLI, scheduled posting
  • Phase 2: Context capture from session history
  • Phase 3: Semantic blueprints (26/26 stories)
  • Phase 4: Brand Planner agent
  • Phase 5: Autonomous content generation
  • Phase 6: Engagement feedback loops
  • Phase 7: Multi-platform (Twitter, blog, YouTube)

Cost

Running the full pipeline on AWS Bedrock: < $1.50/year for 30 posts/month. Local dev with Ollama is free.

License

MIT