Different Hunger

Foundation

Workflows

How execution happens. The autonomous pipeline + role-gated lock flow.

Workflows · How Execution Happens

The autonomous pipeline from intake → live brand domain. Wall-clock targets in minutes. Role-gated review between AI work.

The full chain

T+0:00  Intake submit (37 fields)
        ↓ POST /api/brand-os/intake
        ↓ Insert brand · auto-fire foundation
        ↓ brand.config.intake.poc_email captured

T+0:01  Generate Foundation
        ↓ POST /api/install/generate-foundation
        ↓ 4-step Claude Opus chain (SP + P1→P2→P3→P4)
        ↓ Persist 13-key foundation JSON to brand.config.engine.foundation
        ↓ ~3-4 min

T+0:04  Generate Components (auto-fired after foundation)
        ↓ POST /api/install/generate-components
        ↓ 6 slices from foundation + 8 fresh Claude gens (parallel)
        ↓ Persist 14 content_json blocks to brand.config.engine.components[slug]
        ↓ ~80s

T+0:05  Generate LOFI Doc
        ↓ Google Docs auto-created · 14 component sections · H1 outline nav
        ↓ Store doc_url in brand.config.engine.lofi_doc_url
        ↓ Telegram alert to POC

T+0:05 → T+QA   COPY REVIEW (POC role)
        ↓ POC opens Doc · Suggesting mode
        ↓ Edits + scores per component
        ↓ POST /api/install/sync-edits  (pulls suggestions → content_json patches)
        ↓ POC says "lock copy #NN" per component
        ↓ POST /api/install/lock-component { stage: "copy" }
        ↓ copy_status = copy_locked · copy_locked_at = ISO

T+QA+10  HIFI BUILD (only on copy_locked components)
         ↓ POST /api/install/render-hifi
         ↓ Per component: load locked content_json · render via Master template · brand vars
         ↓ Hosts at /studio/[slug]/install/preview/<component>
         ↓ ~30s

T+QA+10 → T+QA+30  HIFI REVIEW (POC role)
         ↓ POC reviews each rendered page in browser
         ↓ POC says "lock hifi #NN"
         ↓ POST /api/install/lock-component { stage: "hifi" }

T+QA+30  PUBLISH (POC explicit "publish" command only)
         ↓ POST /api/install/publish
         ↓ Auto: DNS pre-flight check (Namecheap getHosts Status=OK · record count ≥20)
         ↓ Vercel alias flip · staging → brand domain
         ↓ brand.config.engine.status = live

Per-component lock gate

draft  →  copy_locked  →  hifi_locked  →  live
        (POC QA Doc)   (POC QA preview)  (POC says publish)

Score ≠ approval. Edits ≠ approval. "Looks good" ≠ approval. Only explicit "lock it" from POC → I call the lock endpoint.

Role routing per action

ActionRole
"lock copy #NN" for brand XPOC of brand X
"lock hifi #NN" for brand XPOC of brand X
"publish" brand XPOC of brand X (+ Auto checks pass)
"approve new master template"Reviewer
"update voice rule for brand X"POC of brand X
"add new system rule"Reviewer
"first-time install for new brand"Reviewer (until proven)
Strategy or architecture disputeFounder (escalation)
Compliance violation (em-dash, fabrication)Auto blocks · Reviewer can override

Sync edits from Doc → Supabase

1. POC edits LOFI Doc in Suggesting mode
2. Each suggestion = textRun with suggestedInsertionIds / suggestedDeletionIds
3. POST /api/install/sync-edits { brand_slug }
4. Parser walks doc paragraphs · tracks current ▸ FIELD context · classifies:
   · score / notes / field_label / content / pure_delete / pure_insert
5. Applies edits to brand.config.engine.components[slug] per field path
6. Field-context fallback: if before-text drifted, overwrite by field path
7. Returns { applied, captured, skipped, report[] }
8. POC reviews report · re-fires Doc gen if needed

Library refs power pattern extraction

At gen time for each fresh-gen component:
1. Pull rated 4-5/5 references from library table (Reviewer-curated)
2. Slug aliases (e.g., landing → landing, landing_page, marketing_website)
3. For each ref: Claude extracts headline_pattern, cta_pattern, structure_pattern,
   voice_pattern, trust_pattern, anti_patterns_observed
4. Inject patterns into prompt as RULES (not verbatim examples)
5. Cache patterns in PATTERN_CACHE (in-memory session)
6. Output reflects patterns applied to current brand's voice

This is why curating refs improves output quality across ALL future gens — pattern extraction compounds.

API surface (callable by AI + humans)

EndpointStatusPurpose
POST /api/brand-os/intake🟢Insert brand row · auto-fire foundation
POST /api/install/generate-foundation🟢4-step Claude chain
POST /api/install/generate-components🟢14 components gen
POST /api/install/sync-edits🟢Doc → Supabase patches
POST /api/install/lock-component🟢Copy or HIFI lock per stage
POST /api/landing/capture🟢Million Verifier · Twilio · Resend
POST /api/install/build-doc🔴Auto LOFI Doc generation (script exists)
POST /api/install/render-hifi🔴Build HIFI from locked copy
POST /api/install/publish🔴Alias flip · DNS check
POST /api/chat/{turn,capture,escalate,poc-reply}🔴AI chat widget backend
POST /api/agent/voice-call🔴VAPI integration

Video Post-Production & Repurposing (SOP)

Canonical SOP (verbatim): Post-Production SOP doc. Transcribe → speaker-attribute → strip filler → correct transcript → timestamp key moments → signature + micro-content.

Timestamp-label legend (from the SOP): EM emotional · QM quotable · TM teachable · @ mention · (blank = chapter header).

MCM caption & name correction glossary (Steve-verified, canonical)

Auto-transcription (whisper/OpusClip) mis-hears these every time. Apply on EVERY MCM clip before publish:

CanonicalNotes
Phyliss Fildermanclient — Steve spells it Phyliss (overrides "Phyllis")
Stan and Ann (Marvers)clients — captions read "Stan and Ann"
EstherPhyliss's caregiver — auto-heard as "Western/Wester"
FatimaStan & Ann's caregiver — auto-heard as "fatigue" / mislabeled "the team"
Christinathe nurse practitioner

Mis-hear → fix: Western/Wester → Esther · fatigue / "the team" → Fatima · nerves → nurse · Yetis → Yiddish · writing → providing · "Doctor Fuss" → the Doctor · "paid" → "recommendations MCM personnel gave to us". Recurring QC: clips often cut off at the end — extend the out-point so the sentence completes. Never ship auto-captions unverified.

Source memory

  • project_role_taxonomy
  • project_install_checklist
  • project_two_stage_lock
  • project_72hr_clients_outputs
  • project_canonical_install_deliverables
  • project_curation_library_triple_play
  • feedback_no_lock_without_mk_qa
  • feedback_no_publish_without_approval
  • feedback_dns_safety_protocol
  • feedback_vercel_promote_alias_gotcha