Back to directory

META v3.1 — LLM Agent Engineering Charter

A compact, always-loaded set of operating rules for frontier coding agents such as Claude Code or Cursor, paired with three explicitly invoked skills for continuous execution, parallel work strands and pre-commitment failure analysis.

META v3.1 — LLM Agent Engineering Charter

META is a short core charter, kept under a hard byte limit enforced by continuous integration, consisting of one guiding bias, one meta-rule and nine numbered rules, loaded automatically every session so a coding agent behaves like a disciplined engineer rather than an eager junior. Three additional skills are invoked only explicitly by the user: zero-pause for continuous execution without artificial pauses, weave for scoped parallel work strands with judgment-based conflict resolution, and premortem for failure analysis before committing to a high-risk plan.

The project's v3 revision removed trigger-phrase auto-activation, where text anywhere in a task, including pasted files or quoted content, could previously switch the agent into an autonomous execution mode; the authors identify this as a prompt-injection surface and now require the skills to be invoked explicitly by the user. An included TypeScript and Docker evaluation harness runs pairwise blind judging and deterministic metrics to compare the charter's effect across versions.

More in Agent skills

See all 581
  • Provenance gate

    github.com

    A deterministic authorization gateway that sits between an AI agent and its tools, tracing where each parameter of a side-effect call came from, and blocking the call or asking a human when the source does not check out.

  • A markdown framework of system rules that restates Isaac Asimov's Four Laws of Robotics as a set of prioritized loss calculations for large-language-model agents to follow, published as a discovery file for agents and crawlers to read.

  • Vitruvius

    github.com

    An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.