tebra-content-os

tebra-content-os

GEO is the new SEO. The zero-click era rewards content engineered for citation, not written and published on hope.

tebra-content-os is a Claude Code-native content operations pipeline built for Tebra's GEO era: content engineered for LLM citation, gated for healthcare compliance at commit time, human-approved on every BOFU asset before publish.

Two delivered artifacts demonstrate the system end-to-end:

• Day 1 — LLM extractability audit (audit/tebra_citation_audit_2026-04-22.pdf): Scored 16 tebra.com URLs across five extractability dimensions. Zero pages passed the 3.5 threshold. Three template-level fixes identified that move 8 of 9 blog posts past threshold in one sprint.
• Day 2 — /brief → /draft pipeline dry run (docs/E2E_DRY_RUN.md): Research brief (briefs/tebra-vs-advancedmd-for-solo-practices.json, 7 proof points, 11 registry-backed sources) produced by brief-author subagent; ~1,400-word BOFU comparison draft (drafts/tebra-vs-advancedmd-for-solo-practices.md) gated through the compliance hook and allowed on first write.

Quickstart

git clone https://github.com/nice-and-precise/tebra-content-os.git
cd tebra-content-os

Environment setup (MCP auth, Python venv, env vars): docs/RUNBOOK.md Section 3.

Day-one demo in Claude Code — paste PROMPT_TO_PASTE_IN_CLAUDE_CODE.md, then:

/audit https://www.tebra.com/features

Returns an LLM extractability score with dimension breakdown and recommendations.

Architecture

Five primitive layers, one responsibility each:

Layer	Location	Role
Skills	.claude/skills/	Reusable cross-tool knowledge: brand voice, healthcare compliance, 4 BOFU asset types, citation block library
Subagents	.claude/agents/	Isolated workers: brief-author, draft-writer, citation-auditor, refresh-auditor, compliance-qa, product-truth, citation-reporter
Hooks	.claude/hooks/	Deterministic gates: compliance pre-check, session context injection, async linting, audit logging
Commands	.claude/commands/	User workflows: /audit, /brief, /draft, /refresh, /citation-report
MCP servers	.mcp.json	External access: Firecrawl (page render) + Exa (search). Post-hire: Search Console, GA4, Asana, HubSpot restore via docs/OPERATING_MODES.md

Decision tree: must block → hook; external data → MCP; multi-step → subagent; reusable knowledge → skill; one-shot workflow → command.

Content pipeline

/brief <query>          # brief-author subagent: retrieval → structured brief
  ↓
/draft <slug>           # draft-writer subagent: brief → compliance-checked draft
  ↓                     # (PreToolUse hook blocks unsourced medical claims)
PMM review              # human approval required before publish
  ↓
Webflow publish         # operator action only (manual in pre-hire mode; see docs/OPERATING_MODES.md)

Compliance model

Every write or edit to drafts/ passes through pre-tool-use-compliance.sh, which calls scripts/compliance_check.py. The runtime hook verifies:

1. Every percentage or clinical-outcome claim detected by the medical-claim regex has a matching citation in the draft frontmatter's sources[].claims_cited[].
2. Every cited source exists in sources/registry.json, is not past its expires_at, and is not authority tier 4.

Denial blocks the write. The hook cannot be bypassed.

approved_for_claims[] is declared on every source in the registry. Validator-time enforcement (python3 -m scripts.validate_drafts) checks alignment for any ClaimCited entry where claim_type is set. Runtime hook enforcement of approved_for_claims[] is a tracked gap — see docs/RESEARCH_GAPS_AND_DECISIONS.md.

Documentation

All specification documents under docs/:

File	Contents
PRD.md	Problem, users, goals, acceptance criteria
ARCHITECTURE.md	Five-primitive wiring using April 2026 Claude Code primitives
DATA_CONTRACTS.md	Pydantic schemas for briefs, drafts, sources, citations, audit events
TASKS.md	Thirteen sequential build milestones with enforced stop points
RUNBOOK.md	Environment setup, MCP auth, troubleshooting, day-two operations
RESEARCH_GAPS_AND_DECISIONS.md	Decision log, open questions, version pinning
REPO_PRINCIPLES_EXTRACTED.md	Design lineage from related production systems

Design principles

1. Git is the database. Every brief, draft, source, and audit event lives in version-controlled files. No separate service, no cloud database.

2. Compliance is architecture. The PreToolUse hook blocks unsourced medical claims at commit time. Policy enforced by the system, not by a checklist.

3. Retrieval before generation. Briefs are built from Search Console, GA4, Firecrawl, and Exa signal before any draft is generated.