Build

Playbook — Automated Content Engine (Indexa Labs)

4 min read

Playbook — Automated Content Engine (Indexa Labs)

Summary

A built-and-running agent system that runs daily, hands-off (“brain-dead”) content for Distribution across TikTok and Instagram (handle: @indexalabs). Agents build a product/market knowledge base, mine it for content angles, produce the posts, and schedule them — governed by three policy layers and shipped via GitHub Actions (“Zernio”). This is a worked Practical Market example of the automated posting + creation + audience-education system we’d discussed in principle.

What it is

An end-to-end content pipeline where agents own the loop from knowledge → angle → asset → schedule → post, with the human only setting policy and reviewing. The niche is research compounds / peptides, so the content is educational (“what is it, really — a 60-second primer, minus the hype”) and must navigate platform censorship.

Architecture

1. Knowledge base (the RAG foundation)

Agents build and maintain a product + market knowledge base — a large, interlinked graph of notes (Bases, Behaviors, Compounds, Concepts, Mechanisms, Methods, Papers, Peptides, Resources, Stacks, Supplements, Transcripts, plus a “Knowledge Base — Strategy & Policy” section). Everything is formatted for RAG retrieval, not human reading.

2. Per-agent account knowledge

Each agent has its own account knowledge folder holding:

  • Brand strategy & tone — how this account speaks.
  • Channel block list — words that can’t be used on that channel (e.g. TikTok keeps deleting the literal compound name, so it’s banned from captions and worked around — “just a short chain of amino acids” instead of the censored term).
  • Account assets: posts, publish, lint, lims.txt, restricted-terms, README, etc.

The daily loop

Runs every day, four agent stages:

  1. Update knowledge — agent scrapes news/sources and updates the knowledge base, formatting for RAG purposes.
  2. Plan — agent mines the knowledge base to produce content angles, lays them into a content calendar, and writes a content brief + updates the tracker.
  3. Produce — agent follows the calendar and creates the content (copywriting + image), saves the asset to its designated location, and marks the tracker complete.
  4. Audit & schedule — agent audits the tracker to find posted vs. not-yet-posted content; anything outstanding gets scheduled for publishing via Zernio (GitHub Actions workflows).
scrape/update KB ──► mine KB → angles → calendar + brief ──► produce copy + image ──► audit tracker ──► schedule via Zernio (GitHub Actions) ──► TikTok / Instagram
   (RAG-formatted)        (tracker)                          (assets saved)         (posted vs not)

The three governing policies

Every agent must follow three policy layers (these are the operating files for this system):

PolicyGovernsExample
Channel policiesPer-platform rulesTikTok vs Instagram formatting; banned/censored words
Brand policiesVoice & styleTone, visual style, education-first framing
Performance policiesWhat to prioritizeLean into content that performs, drop what doesn’t

Channel-specific adaptation (the clever bit)

Same angle, reshaped per platform and per platform’s censorship:

  • TikTok — the compound’s name keeps getting auto-deleted, so posts censor/euphemize it (“[CENSORED] … just a SHORT CHAIN of amino acids”). Photo posts use TikTok’s own settings: optional photo description (≤4000 chars), “save to TikTok inbox as draft”, and let TikTok auto-add recommended music (auto song pick). Observation: TikTok’s algo rewards this native behaviour.
  • Instagram — same idea rendered in a cleaner “research notes / primer” carousel format (“What is it, really? A 60-second primer, minus the hype”).

Scheduling & publishing

A Posts dashboard manages scheduled vs published content. Outstanding items are scheduled straight through, and publishing fires via GitHub Actions workflows (“Zernio”) — no manual posting step.

Why it works (learnings)

  • Knowledge base as the moat. Because angles are mined from a deep, RAG-formatted KB, content is differentiated and defensible rather than generic.
  • Policy layers make it safe to run hands-off. Channel + brand + performance policies are the guardrails that let the human step back. See The approval-gate rule.
  • Censorship is a content format, not a blocker. Treating platform word-deletion as a creative constraint (euphemism + “amino acid” framing) turned a problem into a recognisable post style.
  • Native platform behaviour beats fighting it. Letting TikTok pick the music and using draft-to-inbox leans into the algorithm instead of against it.

Status

Built and running. This is the reference implementation for Distribution automation — copy its structure (KB → per-agent account knowledge → daily loop → 3 policies → Actions scheduling) for new channels.

Graph View

Backlinks