◇ Pipeline · how the workforce runs

One picture. End to end.

Every content piece flows through these 8 stages. HITL gates (gold borders, ◍ lock glyph) are mandatory operator approvals enforced in code, not in agent prompts. The loop arrow at the bottom closes the cycle — engagement signals from published posts seed the next research cycle.

HITL gate (mandatory operator approval)◍lock = HITL enforced in code, not in agent promptDONEACTIVEPENDING(when on a draft's lifecycle)

What each stage does

Click any stage to jump to its operator surface

① Niche ·
agent: operator
Pick which of the 18 niches the next piece serves. Each niche carries audience, primary platform, RPM range, affiliate stack, and headline DNA.
inputs
canon/VIRAL_NICHES_LAUNCH.md strategy doc
outputs
niche.slug → seeds the brief
Open /studio/niches
② Research ·
agent: system + Claude
Pulls top YouTube videos for the topic, classifies their comments (appreciation / suggestion / criticism / gap), and synthesizes a gap brief — what the audience asked for and didn’t get.
inputs
Topic · Region · Platform
outputs
gap_brief → ideate seed
Open /studio/research
③ Ideate ·
agent: agent:brainstorm
Capture raw idea OR consume a research gap. Agent expands into 5–7 concept directions. You pick one (manual mode) or auto-flow takes the top concept.
inputs
raw text / link / transcript / gap brief
outputs
idea → concept(s)
Open /studio/ideate
④ Brief ·
agent: system
Auto-composes a fully-formed agent brief from the concept + niche metadata: audience, platform, brand vibe, cadence context, sample headlines, affiliate opportunities, deliverable instructions.
inputs
concept · niche metadata
outputs
brief row → drafting
Open /api/briefs
⑤ Draft ·
agent: agent:claude / agent:gemini
LLM call against the routed provider (Claude Sonnet for brand-critical, Gemini for bulk). Returns markdown long-form / channel-native social / video script / newsletter copy.
inputs
system prompt + brief.prompt_text
outputs
draft v1 (status: pending_hitl)
Open lib/anthropic + provider router
⑥ HITL gate ·
HITL gate ◍agent: operator
MANDATORY review. Approve (with optional notes) / Reject (with reason) / Regenerate (creates v2). The two-step confirm + audit log entry = code-level gate, not an agent prompt.
inputs
draft v1+
outputs
status → approved | rejected | regenerated
Open /studio/publish
⑦ Adapt ·
agent: agent:claude / agent:gemini
One approved source draft → nine channel-native variants (LinkedIn post, X thread, YouTube script, Newsletter pull-quote, Blog post, Image-gen prompt, Hero image, 15s clip, 60s voiceover). Each adaptation gets ITS OWN HITL preview card.
inputs
approved source draft + target channel
outputs
adapted draft (status: pending_hitl AGAIN)
Open /studio/publish (per-draft action buttons)
⑧ Distribute ·
HITL gate ◍agent: system (postWithGuards)
Passes through 4 security guards (HITL gate → token refresh → rate limit → audit log), then calls the per-platform adapter (LinkedIn / X / YouTube / Instagram / Facebook / TikTok).
inputs
approved adapted draft + channel connection (OAuth)
outputs
published_url + audit log entry
Open /studio/channels

✦ THE STANDING RULES (every stage honors these)

HITL on every public asset — gates 6 and 8 are mandatory; nothing ships without you reading the rendered text.
Per-agent daily cost ceiling — Studio Writer $8 / Social $3 / Video $10 / Newsletter $3 = ~$24/day max workforce burn.
PII-safe audit log — every state transition logged to agent_action_log with whitelisted metadata (never post content).
Tokens encrypted at rest — AES-256-GCM via TOKEN_ENCRYPTION_KEY for every OAuth channel connection.
Persona discipline — clients never see internal agent codenames; public assets are signed by Director personas (per canon).