Fable5 Push Skill

FABLE-5 ULTRA

**Any agent. Any model. Ships like a 20-year engineering org.**

A portable elite-output protocol for **any agentic AI or coding agent** — Claude Code, OpenAI Codex CLI, Cursor, Windsurf, Cline, Roo Code, Amp, GitHub Copilot, JetBrains Junie, Aider, OpenHands, Devin, or any chat API — tuned to the verified behavior of **Claude Fable 5** (Anthropic's Mythos-class model, June 2026) and degrading gracefully on every other model.

The pipeline

For any `fable`-tagged task the agent runs:

0 INTAKE → 1 ARCHITECT → 2 PLAN → 3 BUILD → 4 COUNCIL → 5 RED TEAM → 6 VERIFY → 7 SHIP

• Spec-locked — testable success criteria, binding non-goals, exact acceptance commands
• 8-role council — architect, security, performance, QA, reliability, data, reviewer, docs; findings as [ROLE] P0/P1/P2: … (file:line)
• Red team before ship — 5 production-death questions, defensive framing
• Rule of evidence — nothing claimed as passing without execution; [UNVERIFIED] tags; the Verdict Block names the model and harness that did the work
• Caveman mode — 70–90% token compression (caveman / caveman off)
• ULTRATHINK — highest effort + explicit self-validation (ultrathink)
• Harness-aware — detects full-agentic (H1) / IDE-approval (H2) / chat-only (H3) and degrades the pipeline honestly instead of faking evidence

Repo layout

fable5-ultra/
├── .claude-plugin/
│   └── plugin.json                 ← Claude Code plugin manifest
├── .cursor/rules/
│   └── fable5-ultra.mdc            ← Cursor project rule (alwaysApply)
├── CLAUDE.md                       ← always-on core for Claude Code
├── CURSOR.md                       ← always-on core for Cursor
├── EXAMPLES.md                     ← worked examples + common mistakes
├── README.md                       ← this file
├── README.zh.md                    ← 中文说明
└── skills/
    └── fable5-ultra/               ← the full skill (canonical)
        ├── SKILL.md                ←