Skip to content

Inspirations

Jump to bottom
Yash Kanadhia edited this page Jun 5, 2026 · 3 revisions

Inspirations

Zeref OS stands on the work of many engineers in the AI agent and LLM tooling space. This page documents the direct influences. Full reference: references/v4x-canon/RESEARCH_RESOURCES.md.

Foundations

AGENTS.md standard

What Zeref OS takes: AGENTS.md as the single source of truth; per-harness files are thin stubs. Lean-file discipline (under 250 lines for compliance). Read/write memory separation.

Karpathy paradigm shifts

What Zeref OS takes: Skill discipline (not skill bloat). Boundary-first reads. Wiki-as-canon. Safety-gated extension. Karpathy's "context engineering" framing as a first-class concern.

Canonical AGENTS.md / CLAUDE.md samples studied

What Zeref OS takes: YAML frontmatter conventions. Skill activation triggers. Compatibility with the de facto multi-vendor format.

Community mental models

What Zeref OS takes: PATTERNS.jsonl as a "log as API" — the harness-agnostic event surface. Closed-tool-loop discipline (connectors share filesystem). Eval-driven rule validation (the validator + CI).

Memory and evolution patterns

  • jack60810/claude-evolve — Darwinian memory with EMA rule ratings
  • MEMORY.md read/write separation — community-wide pattern, multiple sources
  • Two-Strikes Rule — derived from Anthropic CLAUDE.md best practices
  • Auto Dream (relative → absolute date hygiene) — community pattern

What Zeref OS takes: Confidence decay + supersession (the contradiction-resolution + archive flow). AGENTS.md = human-written / MEMORY.md = agent-written. Two-Strikes Rule for rule lifecycle. Auto-hygiene on every /done (relative dates → absolute).

Spec-driven development

What Zeref OS takes: Human-gated phase progression (the conversational project-setup interview). Steering-file pattern adapted as harness stubs.

Knowledge management

What Zeref OS takes: Per-domain wiki organization. Boundary-first read protocol.

Key community data points (May 2026)

Metric Value Source
AGENTS.md repos 60,000+ agents.md community
forrestchang/andrej-karpathy-skills installs 43,000 in one week community telemetry
BMAD-METHOD GitHub stars 43,000 github.com/bmad-method
CLAUDE.md instruction compliance ~70% community consensus; red lines need PreToolUse hooks
Optimal CLAUDE.md length 60–250 lines typical projects
Frontier-model reliable instruction count ~150–200 before compliance decays community testing
Context window utilization target 40–60% (Huntley) above 60% = quality drops

Where Zeref OS diverges from community defaults

Community default Zeref OS choice Reason
CLAUDE.md as primary file AGENTS.md primary, CLAUDE.md = stub Harness-agnostic
Skills bundled in harness Skills recommended only User consent + portability
Memory in hosted service Memory in local markdown Privacy-first
Team agents always available On-demand only Token budget
Immediate skill activation Review-first, approval required Prevents misdetection
Single-user persona Free for all users Adoption

How to credit when forking

If you fork Zeref OS or adopt its patterns, attribution to the upstream community is appreciated:

Built on Zeref OS (https://github.com/kanadhiayash/zeref-os) — itself inspired by:
- agents.md standard (Linux Foundation)
- Karpathy paradigm shifts (autoresearch, llm-council, LLM Wiki)
- BMAD-METHOD spec-driven development
- jack60810/claude-evolve memory evolution
- Anthropic CLAUDE.md best practices
- ...and the broader AI engineering community

The full reference list lives in references/v4x-canon/RESEARCH_RESOURCES.md.

Zeref OS

Getting Started

Architecture

Reference

Clone this wiki locally