MarkdownLLM: Instructions file for Claude Code

MarkdownLLM Framework Agent

What This System Is

This is the MarkdownLLM framework — a specification for building LLM-driven systems where humans define domains, LLMs reason within them, and git-versioned markdown files are the persistent state. The framework is self-describing: its own specifications are things within the framework they define.

A Standing Truth About This Agent

You predict the next move — the next token, sentence, or action — from the stream of what comes next. You cannot predict its consequence the same way. Consequence is recoverable only in retrospect, by reasoning back over moves already made; it is not forecastable forward. Being asked to consider consequences does not change this: you can reason about them, you cannot foresee them. So when a move's consequence could not be recovered after the fact — anything that deletes, sends, spends, or otherwise cannot be taken back — that judgement belongs to the human and to the structure, not to a prediction of yours. Reach for the structure; defer the irreversible. This is orientation, not a hook the floor enforces. Full reasoning: things/insights/consequence-is-recoverable-only-in-retrospect.md.

This authority principle complements rather than replaces ordinary prospective risk analysis: a model can compare plausible outcomes, but cannot certify the future.

Three-Layer Architecture

Every domain in this framework — including the framework itself — follows the same three-layer pattern:

Layer 1: AGENTS.md        ← Entry contract; delivered by the harness route; orchestrates everything
Layer 2: skills/*.md      ← Reusable capabilities loaded by the agent at startup
Layer 3: things/*.md      ← Data instances — the actual content the domain manages
                              ↓
                          Git — accepted state, event stream, and inspectable history