A Build Kit to help you build your own version of a content pipeline app using AI coding agents like Claude Code, Codex, or Gemini.
What you'll build
A self-hosted app that runs your whole content operation from one place. You drop ideas into a single inbox, approve the ones worth developing, and the app fans each approved idea into one draft per destination — a short-form social post, a long-form social post, a blog post, whatever destinations you've configured. You write the drafts (or your AI agent does), drag them onto a calendar, and the app publishes them automatically at the scheduled time.
What makes it different is that it's agent-native: external AI agents read your voice training and drive the whole pipeline through a clean REST API. You stay in control — you triage the inbox, you approve the drafts, you review what the agent has learned about your voice — but the agent does the heavy lifting on drafting and rewriting. Every action is logged with the actor (you or the agent) so you can see exactly what happened.
Who it's for
A solo content creator (or a creator working alongside one or more AI agents) who:
- Publishes regularly across multiple destinations — short-form social, long-form social, blog, and so on.
- Wants AI agents to do the drafting, but wants to stay in control of voice, curation, and final approval.
- Wants a single source of truth for ideas, drafts, and scheduled posts instead of stitching together separate tools.
- Is comfortable running their own self-hosted app, or willing to learn.
This is not a hosted SaaS, an agency tool, or a team-with-roles product. It's opinionated about being single-user, single-tenant, and self-hosted.
What it solves
- Ideas slip through the cracks when capture and drafting live in different tools.
- Drop-in AI tools generate generic prose or take over editorial judgment.
- Multi-destination publishing turns into per-destination juggling and copy-paste.
- Voice prompts decay over time, with no review loop for what the agent is learning.
Key user flows
- A user can drop an idea into an inbox in seconds, with a title, a markdown body, and a category.
- A user can approve, reject, or archive any idea — and watch approved ideas fan out into one draft per destination automatically.
- A user (or an AI agent via API) can write each per-destination draft, respecting that destination's character limit and capability flags.
- A user can leave threaded markdown comments on any idea or draft; the agent reads them and steers accordingly.
- A user can drag a ready draft onto a calendar slot, or click "Publish now" with a confirmation.
- The app auto-publishes scheduled drafts at the slot time, captures the live URL, and surfaces failures visibly.
- A user can maintain voice training (general voice, content modes, current interests, per-destination, per-type) from one Settings surface.
- A user can review learnings the agent proposes — voice and style observations — and approve them before they update training.
- A user can see two split todo lists — one for what's waiting on them, one for what's waiting on the agent — both computed live from the data already in the system.
- A user can browse an audit log of every action, tagged with whether it was them or the agent.
How you'll build it
- Idea Inbox — project foundation, single-user auth + per-user API token, the Sparks inbox, spark types, polymorphic comments, the markdown editor, and the Settings shell.
- Flames & Drafts — Platforms with capability flags, approving an idea to fan out per-destination drafts, derived flame state, per-draft comments.
- Schedule & Auto-publish — calendar, manual and simple-recurring slots, drag-and-drop, idempotent publishing with permalink confirmation polling, and the Integrations surface.
- Agent API, Training, & Learnings — the bearer-token JSON API, the Training surface, the Learnings approval lifecycle, the public API docs page, and the agent skill that doubles as the API contract.
- Activity, Todos, & Public Face — the activity feed, live-computed User Todos and Agent Todos, the public marketing homepage, the public user docs page, and final theming polish.
Each milestone is a self-contained, demoable checkpoint. You stop at the end, run what you've built, and move on.
What's included
README.md— orientation and how to use the kit_build_plan/PRD.md— the product requirements, data model, and milestone breakdown_build_plan/tech-stack.md— customizeable tech stack for your codebase_build_plan/design-system.html— visual reference (colors, type, components — open in a browser, no build step)_build_plan/agent-skill.md— a paste-in skill for an AI coding assistant that lets it interact with your running deployment as a content agent_build_plan/milestones/<N>-<slug>/prompt.md— one folder per milestone with a lean, ready-to-paste prompt that works with any AI coding agent
