MemoryAI Coding-Guard

Universal coding-agent guardrails. Install once → 11 IDE/CLI follow the rules.

Stop AI agents from being sloppy. 4 core coding principles + anti-sloppy enforcement — packaged as drop-in rule files for every major coding agent in 2026.

What it does

AI coding agents share the same bad habits: they refactor code you didn't ask them to, over-engineer simple tasks, fake "done" without running tests, and hallucinate APIs that don't exist. Coding-Guard installs typed behavioral rules into your project, plus enforcement hooks that block sloppy actions instead of just suggesting they're bad.

npx @memoryai.dev/coding-guard install

That single command:

• Detects which agent(s) your project uses (Cursor, Claude Code, Aider, Copilot, ...)
• Drops the right rule file into the right path with the right format
• Optionally wires up enforcement hooks (block --no-verify, force test before commit, etc.)

The 4 principles

Distilled from years of senior-engineer practice on LLM coding workflows.

| Principle | What it stops | |---|---| | Think Before Coding | Silent assumptions, missing tradeoffs | | Simplicity First | Over-engineering, speculative abstractions | | Surgical Changes | Drive-by refactors, scope creep | | Goal-Driven Execution | "Done" without verification |

Plus 5 anti-sloppy mechanisms: anti-hallucinate, anti-fake-verify, anti-out-of-scope, anti-loop, anti-destructive.

Every rule ships with rationalization tables (the lies an agent tells itself, paired with the truth) and red flags (observable anti-patterns), wired to runtime hooks that quote the truth back when the lies surface.

Skills, personas, and slash commands (v2.0)

Coding-Guard v2 ships an opinionated workflow layer on top of the rules:

• 7 skills — spec-driven-development, planning-and-task-breakdown, test-driven-development, doubt-driven-development, code-review-and-quality, security-and-hardening, debugging-and-error-recovery. Each is a process with steps, rationalization table, red flags, and verification.
• 4 personas — code-reviewer, security-auditor, test-engineer, web-performance-auditor.
• 8 slash commands — /cg:spec /cg:plan /cg:build /cg:review /cg:audit /cg:debug /cg:doubt /cg:ship (namespaced under cg: to avoid collision).
• 4 reference checklists — security, performance, accessibility, testing patterns (under references/).

Skill / persona / command content is adapted from addyosmani/agent-skills (MIT) and restructured into TOML so the build pipeline can fan it out across 11 IDE/CLI. See NOTICE for attribution. Coding-Guard adds the enforcement layer (predicates + hooks + signed constitution); the upstream playbook supplies the workflow content.

Supported platforms (11)

| Platform | File installed | |---|---| | Claude Code | CLAUDE.md + plugin/hooks | | Cursor | .cursor/rules/*.mdc | | Windsurf | .windsurfrules | | Cline | .clinerules/ | | Roo Code | .roo/rules/ (per-mode) | | Kilo Code | .kilocode/rules/ | | GitHub Copilot | .github/copilot-instructions.md + chatmodes | | JetBrains Junie | .junie/guidelines.md | | Aider | CONVENTIONS.md + .aider.conf.yml | | OpenAI Codex | AGENTS.md + ~/.codex/config.toml | | Gemini CLI | GEMINI.md + ~/.gemini/settings.json |

AGENTS.md is shared across OpenCode, Continue, Crush, and any RFC-9999-compatible agent.

Install

npm (recommended)

npm i -g @memoryai.dev/coding-guard
coding-guard install

Or run directly without installing:

npx @memoryai.dev/coding-guard install

curl (Linux / macOS)

curl -fsSL https://memoryai.dev/install.sh | bash

PowerShell (Windows)

iwr https://memoryai.dev/install.ps1 | iex

Commands

coding-guard install            # detect IDE + install
coding-guard install --all      # install for every platform
coding-guard install --only cursor
coding-guard install --dry-run  # preview without writing

coding-guard list               # list supported platforms
coding-guard sync               # rebuild platform files from source
coding-guard update             # update to latest rule version
coding-guard doctor             # check installation health
coding-guard validate           # validate skills/personas/commands TOML
coding-guard verify             # verify Ed25519 signature of rules

Architecture

rules/
  constitution.toml        ← source-of-truth (typed, signed)
  schema.json              ← JSON Schema for validation
  predicates/*.ts          ← executable assertion functions
  metrics.json             ← per-rule measurable metrics
  lineage.jsonl            ← append-only history (audit chain)
  skills/*.toml            ← 7 workflow skills
  personas/*.toml          ← 4 reusable personas
  commands/*.toml          ← 8 slash commands

platforms/                 ← built outputs per platform (12 dirs)
runtimes/hooks/            ← enforcement hooks (block-dangerous, post-edit-verify,
                              stop-gate, inject-rules, quote-rationalization,
                              detect-red-flags)
references/                ← security / performance / accessibility / testing
telemetry/                 ← optional opt-in violation logging
scripts/sync-upstream.js   ← upstream diff report for vendored skills

The constitution is signed Ed25519 from day one. Future-proof for cross-vendor rule registries (RFC-9999 AGENTS.md, Sigstore Rekor).

Family

@memoryai.dev/coding-guard is part of the MemoryAI Guards family:

• @memoryai.dev/context-guard — protect AI context from overflow
• @memoryai.dev/coding-guard — this package
• More guards coming.

License

MIT

Built by MemoryAI. For developers who want their AI agents to think before they type.