META Engineering Charter
Give coding agents compact rules for sound judgment, verification, and irreversible actions.
- Source repo
- entropyvortex/meta-llm-charter
- Stars
- ★ 247
- Last updated
- 6d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 65/100 · Some gaps
At a glance
- How it runs
- Works with
- Portable with changesClaude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An engineering team making consequential code changes installs the charter in a Claude Code project to require reproducing failures and verifying fixes.
- Not a fit if
- Teams that require published v3 evaluation results
- Teams that need the skills to work directly outside Claude Code
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-14
- Source review
- 65/100 · Some gaps
What does this agent do, and when should you use it?
META v3.1 is an operational charter for coding agents, centered on CLAUDE.md, which loads each session and contains a Bias, the META-0 meta-rule, and nine rules. The repository also provides three explicitly invoked Claude Code skills: /zero-pause, /weave, and /premortem, plus an optional Weave scope-guard.sh write-boundary hook. The core rules cover task decomposition, asking, reproducing failures, test contracts, evidence tags, dissent, and confirmation gates for irreversible actions. The core can be used with Claude Code, Cursor, or other agents; the skills and hook are Claude Code-native. The README says it is used in production work at Vortex Serviços and SabIA, but no v3 evaluation runs have been published.
For setup, users can place CLAUDE.md in a project for Claude Code to read each session, paste its contents into Cursor Rules, or provide it to another agent as a high-priority system prompt. The agent follows rules such as decomposing work and identifying root causes and invariants before coding, and reproducing failures before repair; R5 and R8 call for evidence tags, while R10 gates irreversible actions such as schema, production-data, and breaking API changes. Users can explicitly invoke /zero-pause for continuous execution, /weave for parallel-strand orchestration, or /premortem for analysis of high-impact plans; scope-guard.sh can be registered as a PreToolUse hook to block out-of-scope writes. The evals/ directory contains a TypeScript and Docker A/B harness for development use, and is not auto-loaded.
- An engineering team making consequential code changes installs the charter in a Claude Code project to require reproducing failures and verifying fixes.
- A developer using Cursor pastes the core rules into Cursor Rules to shape task decomposition and testing behavior.
- A team splitting work into independent strands explicitly invokes /weave in Claude Code to coordinate parallel work.
- An engineering lead preparing an architectural commitment or high-impact launch invokes /premortem to examine fragile assumptions.
- A developer seeking continuous momentum on a well-scoped task invokes /zero-pause and uses the humanpending.md protocol for items that truly require human input.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Config files
.claude/settings.json
How do you install or deploy this agent?
For the minimal installation, download the core file:
curl -O https://raw.githubusercontent.com/entropyvortex/meta-llm-charter/main/CLAUDE.mdFor the full installation, clone the repository and copy the core file and skills:
git clone https://github.com/entropyvortex/meta-llm-charter
cp meta-llm-charter/CLAUDE.md your-project/
mkdir -p your-project/.claude
cp -r meta-llm-charter/.claude/skills your-project/.claude/skillsTo use the optional scope guard, register .claude/skills/weave/hooks/scope-guard.sh as a PreToolUse hook in the project’s .claude/settings.json. The README does not provide the hook’s configuration contents.
How do you use this agent?
Claude Code reads CLAUDE.md automatically from the project; after installing the skills, invoke one when needed:
/zero-pause
/weave
/premortemCursor users can paste CLAUDE.md into Cursor Rules (or .cursor/rules). For other agents, provide the core as a high-priority system prompt; to use a skill mode, include its SKILL.md in context. Skills do not activate from trigger phrases in task files, quoted text, or tool output.
What are this agent's strengths and limitations?
- The 2,399-byte core can stay loaded in every session, with a CI size gate set at 2,400 bytes.
- R10 defines explicit confirmation gates for actions such as schema changes, production-data mutation, and breaking API changes.
- Three opt-in skills cover continuous execution, parallel-strand orchestration, and pre-mortem analysis of high-impact plans.
- The optional scope-guard.sh hook can deterministically block writes outside a claimed Weave strand.
- No v3 evaluation runs have been published, so the current design has no public run results validating its performance.
- The skills and scope hook are Claude Code-native; other agents get the core or must be given a SKILL.md manually.
- The charter may over-caution on fuzzy or exploratory work; the README suggests explicitly invoking /zero-pause where appropriate.
- The README describes a single-author origin and says performance varies by base model.
How does this agent compare with similar options?
The README contrasts v3 with v2: v3 removed trigger-phrase auto-activation and requires explicit user invocation of skills. It positions META as a charter for coding agents such as Claude Code and Cursor, and traces its foundational principles to forrestchang/andrej-karpathy-skills.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| META Engineering Charter This agent | 65 · Some gaps | Agent plugin / skillFree | ★ 247 | 6d ago | TypeScript | Claude Code |
| HERO — Anti-OverDefense | 64 · Some gaps | Agent plugin / skillFree | ★ 886 | 3d ago | Python | Codex · Claude Code |
| Loop Engineering | 65 · Some gaps | CLIFree + model costs | ★ 12k | 1d ago | TypeScript | Codex · Claude Code |
| CoCo Super Intelligence | 64 · Some gaps | Agent plugin / skillFree + model costs | ★ 327 | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The charter constrains agents with minimal-fix estimates, scope budgets, and explicit confirmation gates. R10 names irreversible actions and limits authorization to its scope, supporting a score of 3 for user confirmation. Deductions: the harness permission system is mentioned but its boundaries are not specified; the files provide limited detail on sensitive data, data flows, dependency security, or rollback. R8 attribution rules and repository authorship offer some provenance, but publisher identity remains unknown.
The rules cover decomposition, reproduction, test contracts, evidence tags, and completion checks, and they generally reinforce one another. They also specify how to record a fork or stop work when a gate applies. Deductions: the sample does not show dependency pinning or availability measures, and failure handling is described mostly as policy rather than with a fuller troubleshooting guide.
The README names suitable engineering work and less suitable exploration or rapid prototyping. The core is presented for several agents, while skills require explicit invocation. Deductions: the skills are primarily Claude Code-specific, cross-environment adaptation guidance is limited, and the charter acknowledges that highly ambiguous requirements can overwhelm a system prompt.
The README clearly separates the core, skills, optional hook, evaluation tooling, and installation paths. The MIT license and stated limitations are clear. Deductions: the sample does not show a full changelog, FAQ, or clearly assigned maintenance responsibility; examples are mainly install commands and protocol formats.
The compact core, irreversible-action gates, evidence tags, and optional zero-pause and parallel-work skills provide a usable workflow for coding tasks. Deductions: v3 has no published evaluation runs, so the current version's benefits have not been measured; the charter may add caution and process overhead to exploratory work.
The charter requires executed, inspected, or assumed tags on decision-facing claims, explicitly labels historical evaluation numbers unverified, and states that v3 has no published runs. The README also describes the evaluation fixtures and run manifests. Deductions: the core and skill files in this sample are not independent evidence of effectiveness, and some README claims about adoption and CI are not fully corroborated by the supplied material.
- The README says v3 has no published evaluation runs; do not treat design goals or historical v2 numbers as verified results for this version.
- The core is presented as portable, but the supplied skills and scope-guard hook target Claude Code. Confirm invocation and permission behavior before adapting them to another environment.
FAQ
Does using the charter cost money?
Can Cursor and other agents use all of the skills?
Do published evaluations show that META v3 improves on default behavior?
Will a skill activate when related phrases appear in a prompt?
Is the scope guard enabled by default?
.claude/settings.json.Is META Engineering Charter open source?
Is META Engineering Charter safe to use?
Is META Engineering Charter still maintained?
Who makes META Engineering Charter?
Thanks — we use this to improve the page.