| name | wr3-editorial-bench |
| description | WR3 editorial-bench skill cortex — operational heuristics, on-tone examples, lessons growth surface. Loaded by wr3-editorial-bench agent at every dispatch. |
| model | opus |
| lifecycle_tier | scheduled |
| cost_class | reasoning |
| ceiling | $0.5 |
| contract_version | 1.0.0 |
WR3 editorial-bench — Skill Cortex
Role (one-line)
Monthly cron 1st Monday 07:00 WITA. Researches SOTA video editorial from 12 reference brands (NYT Opinion Video, Bloomberg Originals, Reuters, Wired, Pudding, Rest of World, ProPublica, Vox, Quartz, Drift, Vice News, Al Jazeera English) + 3 Bali Zero competitor video accounts (Lets Move Indonesia, Emerhub, Flado IG Reels). Multi-LLM: Gemini 3.1 Pro for long-context ingestion, Claude Opus for synthesis, DeepSeek for pattern extraction.
Primary I/O
- Inputs: 12 reference brand video archives + 3 competitor archives + 2 trend reports
- Outputs: ~/.claude/skills/bali-zero-brand/_external-bench-video-YYYY-MM.md
Symbiosis law emphasis
Law 8 (Passato/Presente/Futuro) — read by yt-metrics-analyst (weekly) and critic (every run via skill load)
On-tone examples
(seeded empty — populated by wr3-reflexion-synth weekly after first 3 episodes pass critic gate)
TBD-2026-05-18 — first 3 pilot episodes feed examples here.
Lessons (operational heuristics)
(growth surface — wr3-reflexion-synth appends max 10 lessons/week here)
2026-05-18 — Foundation seeded
Skill cortex created at S7.3 step of WR3 genesis. No operational lessons yet.
Anti-patterns (banned)
- (seeded empty — anti-patterns added as critic gate FAILS produce lessons)
Output structure (_external-bench-video-YYYY-MM.md)
- Executive Summary — Claude Opus 4.7, 200-300 words, BZ-field framing
- Trends Du Mois — agy long-context, 10-15 bullets rising/falling
- Numerical Patterns — DeepSeek V4 Pro, p50/p90 (hook ms, cuts/sec, text density chars/sec, emoji freq, title length)
- Bali Zero Adoptable Patterns — Claude synthesis, 5 testable bullets
- Anti-Patterns to Avoid — Claude synthesis, 5 bullets
- Per-Brand Briefs — agy, ~100 words / brand × 15
- Cost Report — per-tier telemetry table (wall, tokens, $)
Cost budget (monthly)
| Tier | LLM | Cost | Quota source |
|---|
| 2 | Gemini 3.1 Pro via agy CLI | $0.00 | Google AI Ultra subscription |
| 3 | Claude Opus 4.7 via claude-cascade.sh | ~$0.30 | Claude MAX OAuth (no SDK) |
| 4 | DeepSeek V4 Pro reasoning_effort=high | ~$0.05 | DeepSeek API (Keychain key) |
| Total | | ~$0.35 | within $0.50 ceiling |
Cost ceiling discipline
ceiling_usd: $0.5 (declared in docs/wr3/contracts/editorial-bench.yaml, hard_halt_on_exceed: true)BudgetExceededError → cascade decision per Law 7 (Numeri prima):
- If agent is GATE (design-architect, pre-render-gatekeeper) → HARD HALT + Telegram P0
- If agent is HOT PATH (lifecycle_tier=core) → cascade to Gemini fallback
- Otherwise → mark FAIL, retry next cycle
Resources
- Agent definition:
~/.claude/agents/wr3-editorial-bench.md
- I/O contract:
~/Desktop/nuzantara/docs/wr3/contracts/editorial-bench.yaml
- Brand cortex (shared):
~/.claude/skills/bali-zero-brand/
- Symbiosis precedence:
~/Desktop/nuzantara/docs/wr3/symbiosis-precedence.md
