PRD: Automatos Autonomous Content Engine

Introduction

Turn the blog backend (built in ralph/blog-widget-backend) into an autonomous content pipeline. Agents research, write, review, and design cover art for blog posts on a schedule. Humans approve with a single click. The blog gets its own top-level nav item in the dashboard, and a new platform_update_blog_post tool lets agents edit drafts collaboratively. This PRD keeps storage simple (PostgreSQL only for v1) and defers social distribution to a future PRD.

Goals

  • Ship a complete agent-driven blog pipeline: write, review, design, queue for approval

  • Add platform_update_blog_post tool so agents can collaborate on drafts

  • Move Blog management out of Activity into its own top-level nav route

  • Provide a sample playbook template ("Weekly Blog Pipeline") in the marketplace

  • Define the QUILL (writer), EDITOR (reviewer), and CANVAS (designer) agent roles

  • Human publishes with one click in the existing Blog Editor — no new approval UI needed

User Stories

US-001: Add platform_update_blog_post tool

Description: As an agent, I need to update an existing draft blog post so that multiple agents can collaborate on the same article (e.g. editor improves what writer drafted).

Acceptance Criteria:

US-002: Add platform_get_blog_post tool

Description: As an agent, I need to read the full content of a blog post so I can review or edit it.

Acceptance Criteria:

US-003: Move Blog to top-level nav route

Description: As a workspace owner, I need Blog as its own page in the main sidebar navigation so it's easy to find and manage separately from Activity.

Acceptance Criteria:

US-004: Create QUILL agent template

Description: As a workspace owner, I need a pre-configured "QUILL" writer agent template so I can deploy a blog writer without manual setup.

Acceptance Criteria:

US-005: Create EDITOR agent template

Description: As a workspace owner, I need an "EDITOR" reviewer agent that reads drafts and improves them.

Acceptance Criteria:

US-006: Create CANVAS agent template

Description: As a workspace owner, I need a "CANVAS" designer agent that generates cover images for blog posts.

Acceptance Criteria:

US-007: Create "Weekly Blog Pipeline" playbook template

Description: As a workspace owner, I need a ready-made playbook that orchestrates the full blog pipeline so I can enable autonomous publishing with one click.

Acceptance Criteria:

US-008: Task-to-publish integration

Description: As a workspace owner, when I complete a "Review & Publish" board task, I need a clear path to publish the blog post without hunting for it.

Acceptance Criteria:

Description: As a user clicking a task link, I need /blog?edit={post_id} to open the editor directly.

Acceptance Criteria:

Functional Requirements

  • FR-1: platform_update_blog_post tool updates specified fields on a draft post, workspace-scoped, returns updated metadata

  • FR-2: platform_get_blog_post tool returns full post content by ID or slug, workspace-scoped

  • FR-3: Blog management lives at /blog as a top-level route with its own sidebar nav entry

  • FR-4: Agent templates (QUILL, EDITOR, CANVAS) are JSON files that can be installed from the marketplace or applied manually

  • FR-5: The "Weekly Blog Pipeline" playbook template orchestrates 4 steps sequentially with agent handoff

  • FR-6: Step 4 of the pipeline creates a Board task with a deep-link to the blog editor

  • FR-7: The blog editor auto-opens when navigated to with ?edit={post_id} query param

  • FR-8: Human publishes by clicking the existing "Publish" button in the Blog Editor — no new approval UI

  • FR-9: Blog posts remain in PostgreSQL only (no S3, no vectorisation in v1)

  • FR-10: Each agent template includes tool assignments so agents are ready to work after installation

Non-Goals

  • No S3 storage or vectorisation of blog posts (future PRD — when we want blogs as RAG-searchable knowledge)

  • No social distribution (LinkedIn, Reddit, X) — separate future PRD

  • No auto-publish on task completion (human clicks Publish manually in v1)

  • No image upload/hosting (agents provide external URLs for cover images)

  • No editorial workflow UI (no approval queue, no diff view between drafts)

  • No RSS/Atom feed generation

  • No blog analytics beyond the existing view_count field

  • No automatic topic selection (QUILL's prompt includes the category, but topic choice is up to the agent)

Design Considerations

  • Blog page reuses the existing ActivityBlog component — just re-mounted at a new route

  • Agent templates follow whatever pattern existing marketplace templates use (check orchestrator/modules/agents/templates/ or marketplace seed data)

  • Playbook template follows the existing playbook/recipe structure with sequential steps

  • The pipeline is opinionated but configurable: users can remove steps, change agents, adjust the schedule

  • CANVAS agent is best-effort — if no image generation tool is connected, it gracefully degrades

Technical Considerations

  • platform_update_blog_post handler must only update fields that are explicitly provided (use **kwargs pattern from BlogService.update_post)

  • Playbook Step 4 needs to create a Board task programmatically — verify platform_create_task tool exists and supports description with markdown/links

  • Agent templates need to reference tool names that exist in the ActionRegistry — verify all tools are registered before shipping templates

  • The ?edit={post_id} deep-link pattern is consistent with how other pages handle deep-links (e.g. Activity page uses ?tab= and ?task_id=)

  • Blog nav item should respect the same permission model as other nav items (workspace member access)

Success Metrics

  • End-to-end pipeline runs autonomously: QUILL drafts, EDITOR reviews, CANVAS designs, task created — with zero human intervention until the publish step

  • Human can go from Board task notification to published post in under 30 seconds (click task link, review, click Publish)

  • Pipeline produces 2 blog posts per week when scheduled

  • At least 1 complete pipeline run verified on automatos.app within first week of deployment

Open Questions

  • Should QUILL use web search (Composio) for topic research, or only workspace memory/RAG? (Recommendation: both, if web search tool is available)

  • Should the pipeline playbook be auto-installed for new workspaces, or only available in the marketplace? (Recommendation: marketplace, not auto-installed)

  • When we add S3 + vectorisation in v2, should we backfill existing posts? (Recommendation: yes, one-time migration)

  • Should the Board task auto-complete when the post is published? (Would need a webhook/trigger on status change — defer to v2)

Last updated