Claude Mythos Scaffold

Pattern-based scaffold for systematic AI agent work, with honest capability boundaries.
Inspired by observed Claude Mythos Preview behaviors (Anthropic, April 2026). Not affiliated with Anthropic.

A Claude Code skill set that brings agentic discipline (Plan-and-Execute, Reflexion, Self-Refine, Ralph Loop) into structured, reusable protocols. Real-world tested in production sessions; capability limits documented honestly.

Why This Exists

Claude Mythos Preview demonstrated agentic behaviors (problem persistence, multi-step iteration, verification-driven correction) that scaffold patterns can partially reproduce in Opus-class and smaller models (Opus 4.8, Sonnet 5, Haiku 4.5). This repo packages those patterns as opt-in skills. Since July 2026 the scaffold also carries core/fable-distilled.md: working patterns written by Claude Fable 5 (the GA'd Mythos-class model) about its own decomposition, verification, and next-action habits, captured during the 2-7 July free window.

Honest expectation: the scaffold transfers process discipline, not capability. It makes Opus 4.8 a disciplined Opus 4.8, not a Fable. Raw reasoning depth, novel pattern recognition, and sample efficiency cannot be patched in via prompts; those are weights. Do not fully load it on Fable 5 / Mythos 5: over-scaffolding degrades Mythos-class output.

What you get: systematic protocols for tool selection, context priming, problem decomposition, verification loops, and failure recovery. Plus two domain modes: long-horizon research synthesis and codebase migration.

Repository Layout

claude-mythos-scaffold/
├── SKILL.md                       Skill entry: model gating + tiered activation
├── core/                          Foundation skills (vault-agnostic)
│   ├── fable-distilled.md         Fable 5 working patterns (read first)
│   ├── mode.md                    Entry point, mode rules
│   ├── tool-stack.md              Cascade selection, parallel/sequential
│   ├── context-priming.md         Adaptive RAG, source hierarchy
│   ├── decomposition.md           Sub-agent delegation, hub-and-spoke
│   ├── agent-loop.md              Plan-Execute, Reflexion, Self-Refine
│   ├── verification.md            Headless verify, output reading
│   ├── failure-recovery.md        Ralph Loop, persistence threshold
│   └── memory.md                  MemPalace integration, AAAK format
├── domains/
│   ├── research/                  Long-horizon multi-source synthesis
│   │   ├── mode.md
│   │   ├── retrieval.md           6-tier source hierarchy
│   │   ├── synthesis.md           Claim graph, contradiction tree
│   │   ├── cite-verify.md         Hallucination check, Feynman pattern
│   │   └── output.md              Academic / blog / brief / slide / wiki
│   └── migration/                 Codebase migration / framework upgrade
│       ├── mode.md
│       ├── audit.md               Footprint, breaking change matrix
│       ├── plan.md                Phasing, calibrated time estimates
│       ├── execute.md             Atomic commits, sub-agent parallel
│       └── rollback.md            4 strategies, drill protocol
├── commands/
│   └── mythos-mode.md             /mythos-mode slash command
├── hooks/
│   └── mythos-sync.py             PostToolUse hook, vault to global sync
└── examples/
    └── (real session walkthroughs, coming)

Quick Start

Option A: Drop into Claude Code

Copy core skills to your global Claude Code skills directory.

macOS or Linux:

cp -R . ~/.claude/skills/mythos-scaffold
cp commands/mythos-mode.md ~/.claude/commands/

Copy the whole directory (not just core/*): SKILL.md resolves skills via relative ./core/ links and drives the tiered activation.

Windows:

Copy-Item -Recurse . $env:USERPROFILE\.claude\skills\mythos-scaffold
Copy-Item commands\mythos-mode.md $env:USERPROFILE\.claude\commands\

Then in any Claude Code session:

/mythos-mode <your task>

Option B: Domain-only

Just need research synthesis or codebase migration scaffold? Copy the relevant domains/<name>/ directory plus core/ (foundation skills are required).

Option C: Vault integration (Obsidian-style)

If you maintain a knowledge vault, the skills support [[wikilink]] cross-references and integrate with MemPalace for verbatim conversation history. See core/memory.md.

What's Tested vs Newer

Production-tested (used across multiple real Claude Code sessions):

• core/mode.md, tool-stack.md, context-priming.md, decomposition.md, agent-loop.md, verification.md, failure-recovery.md

Newer (1 internal review pass):

• core/fable-distilled.md (distilled from Claude Fable 5 itself, 2026-07-06)
• core/memory.md (MemPalace integration)
• domains/research/* (5 skills)
• domains/migration/* (5 skills)

Real-world feedback expected; PRs welcome.

Acknowledgments

This scaffold is a synthesis of patterns from active open-source projects, research papers, and industry case studies. Full credit list with links: see REFERENCES.md.

Highlights:

• Claude Mythos Preview (Anthropic, April 2026): the observed behaviors that motivated this scaffold
• MemPalace (Jovovich + Sigman, 2026): AAAK compression, hierarchical loci memory, used in core/memory.md
• claude-mem (thedotmack): Claude Code-native memory plugin alternative
• Ralph Loop (Geoffrey Huntley): persistence pattern, used in core/failure-recovery.md
• Self-Refine (Madaan et al.): iterative output refinement, used in core/agent-loop.md
• Reflexion (Shinn et al.): episodic self-critique
• ReAct (Yao et al.): think-act-observe loop
• Adaptive RAG (Jeong et al.): query-complexity routing, used in domains/research/retrieval.md
• PaperOrchestra (Google Research): multi-agent research synthesis
• Environment-in-the-Loop (ReCode 2026): code migration with environment integration
• Aviator Java to TypeScript case study, Doctolib production migration (2026 enterprise reports)

This repo is independent. Not endorsed by Anthropic, MemPalace, or any cited project. Patterns are credited; opinions and integration choices are mine.

Contributing

PRs welcome, see CONTRIBUTING.md. Particular areas of interest:

• Cross-platform sync hook (currently Windows-tested, Unix variant needed)
• Real-world session case studies (anonymized)
• Generic versions of vault-specific references (some skills mention conventions like (C) prefix; these are opt-in)
• Additional domain modes (data engineering, devops incident response, content writing)

License

MIT, see LICENSE.

Anti-Goals (What This Is Not)

• "10x your AI agent" magic. Scaffolding helps consistency, not raw capability.
• Drop-in replacement for engineering judgment. The scaffold structures decisions; humans still make them.
• Production-stable v1.0. This is v0.1; patterns are sound, edge cases will surface.

If you're looking for those things, this repo will disappoint. If you're looking for systematic AI agent discipline with honest documentation, it's a starting point.