Animus/ Blog
runtime ↗animus.team ↗github
← All posts
DOGFOODING · 6 min read · The Animus Teamcontent-opsagent-orchestration

How AO writes its own blog posts

Most AI content tools are black boxes that output generic posts and pray you do not look under the hood. We built the opposite. The animus-blog-generator pipeline is open source, runs on a cron schedule, and commits finished markdown directly to GitHub. You can read every prompt, every agent definition, and every handoff contract in the launchapp-dev/animus-blog-generator repo. Here is exactly how it works, from topic discovery to committed draft, and what we learned about letting agents ship content without a human babysitter.

The Two Ways a Post Starts

Every post enters the system through one of two doors.

The first is a cron schedule. By default, the pipeline tries to ship twice a week: Tuesday at 8 a.m. and Thursday at 8 a.m. The blog-production workflow wakes up, picks its own topic, gathers sources, writes the draft, optimizes it, generates a featured image, writes social excerpts, and pushes to origin. If you configure Supabase, it also upserts the post to a database. If you do not, git remains the source of truth and the phase skips cleanly.

The second door is a human-approved Linear ticket. The blog-from-ticket workflow checks the ticket status, converts the ticket body into a brief, then runs the exact same content pipeline. When it finishes, it posts a completion comment and optionally moves the ticket to done. A cancellation guard re-checks Linear status before every phase, so a human can kill the job mid-flight.

Ten Phases from Idea to Commit

The scheduled pipeline is relentless. Here is the full chain:

  1. Topic research. A content-strategist agent scans Search Console for striking-distance keywords, checks competitor blogs for gaps, and mines Reddit and forums for real questions people are asking.
  2. Research collection. A content-researcher gathers citable sources using Firecrawl, Exa, Tavily, Brave, and Google Maps.
  3. Content writing. A content-writer running Claude Opus 4.6 writes content/<slug>.md with full YAML frontmatter: title, slug, keywords, schema markup, word count, and more.
  4. Commit draft. A bash phase git-commits the file.
  5. SEO review. An seo-optimizer audits keyword density (target 0.5 to 1.5 percent), fixes meta descriptions, adds internal links, pushes readability to grade 8 to 10, and scrubs AI clichés.
  6. Asset generation. An asset-generator calls Replicate to generate a 16:9 WebP featured image using the google/nano-banana-pro model.
  7. Social excerpts. The same agent writes platform-specific social copy to <slug>.social.md.
  8. Register post. A register-post-runner appends the post to content/manifest.json for deduplication and internal-link indexing.
  9. Push branch. A bash phase pushes to origin.
  10. Publish post. An optional command-phase upserts the post into Supabase/PostgREST. If unconfigured, it skips without error.

The ticket-driven variant reuses phases 3 through 10 verbatim after converting the human brief.

Who Does the Work

The pipeline is not a single model with a long prompt. It is a chain of specialized agents, each with its own model, MCP servers, and mission.

  • content-strategist: Claude Sonnet 4.6, armed with Search Console, Exa, Tavily, Brave, Firecrawl, and a content library. Picks the topic and writes the brief.
  • content-researcher: Claude Sonnet 4.6, same research stack. Gathers the raw material.
  • content-writer: Claude Opus 4.6, the only Opus in the chain. Writes the long-form draft.
  • seo-optimizer: Claude Sonnet 4.6. Audits and fixes SEO in place.
  • asset-generator: Claude Sonnet 4.6. Handles images and social copy.
  • performance-analyst: Claude Sonnet 4.6. Runs weekly refresh analysis.
  • approval-watcher: Claude Haiku 4.5. Polls Linear every 15 minutes for human approvals. This is a deterministic shell script, not an LLM call, so deduplication is exact and the polling costs zero tokens.

There is also an optional discovery loop. Every morning at 7 a.m., a transcript-collector and idea-strategist ingest meeting transcripts, propose 3 to 5 blog angles, validate them against SEO and competitor data, and create Linear issues at status ready. A human edits, approves, and the ticket-driven pipeline takes over.

How We Keep Quality from Drifting

Handing a content pipeline to agents is only safe if the guardrails are explicit.

The central contract is business-context.yaml. Every content agent reads it before writing. It defines the niche, the audience, the tone, banned words, and competitive differentiators. If an agent drifts, it drifts against a document it literally just ingested.

Best practices live in .animus/skills/: content-strategy.md, ai-seo.md, content-humanizer.md, seo-audit.md, schema-markup.md, and social-content.md. These are not loose guidelines. They are skill files the agent loads as part of its context.

Phase handoffs are enforced by output contracts. If content-writing omits the slug field, register-post cannot run and the workflow fails fast. There is no implicit ambient state. Every downstream dependency is declared.

A content-library MCP server acts as a bring-your-own knowledge brain for brand positions and established facts. If it is absent, agents fall back to content/manifest.json plus external research. The library and transcript integrations ship as no-op command: "true" stubs so the YAML compiles before any secrets are configured.

Lessons from Shipping Without a Human in the Room

We have been running this long enough to earn a few scars. These are the design decisions that matter:

  • Schedules ship disabled by default. Every cron trigger starts with enabled: false. You must explicitly flip it after loading your .env. This prevents a fresh clone from burning through API credits before you know what happened.
  • Git is the source of truth. content/manifest.json is checked in. Database publishing is an opt-in downstream target. This makes the pipeline auditable, offline-capable, and trivial to roll back.
  • You must compile workflows to disk. Any change to a YAML file must be followed by animus workflow config compile. The daemon compiles in-memory on startup, but the runner needs the persisted compiled config. Skip this step and you get "no workflow found for subject."
  • Human gating happens outside the LLM. The approval watcher is a shell script, not an agent conversation. That keeps it deterministic and free.
  • The daemon does not auto-load .env. You must source .env into the daemon's shell. This prevents secrets from leaking into compiled workflow JSON.
  • Asset generation is currently a hard requirement in the scheduled pipeline. If you lack a Replicate token, you remove the phase rather than expecting an automatic skip.

What You Should Do Next

If you run a technical blog and you are tired of copy-pasting between research tabs, writing the same meta-description boilerplate, and forgetting to update your internal link index, this pipeline is built for you. Clone launchapp-dev/animus-blog-generator, read the YAML, fill in your .env, and compile the workflows. The hardest part is deciding what you want to say. The agents can handle the rest.

Sources