The most expensive bug in AI-assisted building is shipping the wrong thing, well. A drop-in CLAUDE.md / AGENTS.md for PM + eng teams using Claude Code, Cursor, or Codex: 7 principles for problem framing, scope, tradeoffs, and outcome measurement. Built on @karpathy's observations.

Enables agents to generate, refine, save, validate, and read sanitized SVG assets from text prompts through a local OpenAI-compatible model, with tools for listing and managing SVG files.

Enables AI agents to build, validate, query, and render 3D homes and object scenes from a single JSON scene file over MCP, with browser walkthroughs for people.

Enables MCP clients to review Ceramic and ordinary websites with desktop/mobile browser QA, deterministic design and accessibility checks, optional visual judgement, and repeatable fix verification backed by preserved screenshots and reports.

Enables searching sticker collections, viewing sticker images for visual models, and sending stickers—optionally with a short caption—to an authorized QQ session through a running QQ bridge.

Enables MCP clients such as ChatGPT, Claude or Grok to search and read the user's own WhatsApp message history through a linked device. Messages are stored encrypted at rest, and access is read-only unless the optional send tool is enabled.

Enables MCP-capable coding agents to act as a deterministic technical art director and multi-engine pipeline builder, locking AI-generated 2D/2.5D game assets to consistent camera angles, palettes, rigs, atlases, and exports across Blender, ComfyUI, UniRig, LibreSprite, Godot 4, Unity, and Unreal Engine 5.

Enables AI assistants to read and write macOS Reminders, Calendar events and Notes through fixed AppleScript tooling, restricted to the lists, calendars and folders you explicitly allowlist. Writes and deletes are opt-in, and no data leaves the app beyond what the allowlist exposes.

Stores what AI agents learn — thoughts, decisions, session events and repo facts — in PostgreSQL with pgvector and serves it back through versioned MCP tools with hybrid vector-plus-full-text retrieval and budgeted context packs. Enables multiple agents to share one scoped, role-isolated long-term memory, with deliberate promotion of distilled candidates into durable memories.

Exposes the battle.training agent tank arena API as MCP tools, letting clients read ladder standings, bot ratings, recent matches and replay URLs, and inspect cup info. With an API key it also enables registering bots, uploading smoke-tested controller versions, and enrolling in sponsored cups.

Enables AI to create and revise kSQL Action plugin configurations from natural-language requirements, using the plugin’s bundled authoring API to validate SQL, APPLY behavior, and formulas while only reading kintone app definitions.

Enables agents to capture real browser interactions over CDP, compile them into deterministic replay scripts with parameterized inputs and stored secrets, and execute those workflows with real input events — no page re-reading or LLM in the replay loop. Also emits portable Agent Skills from recorded workflows.

Enables Claude to read recent activities and wellness data (sleep, HRV, resting heart rate) and to create, update, or delete structured planned workouts through intervals.icu, which syncs them to Garmin Connect and onto the user's watch.

Exposes local recommendation tools over MCP so clients can rank groceries with explainable scoring across nutrition, price, promotions and close-to-expiry offers. Also supports configurable offer monitoring, weekly plan digests and deduplicated notifications backed by SQLite state.

Enables AI coding agents to save, recall, read, list, and forget plain Markdown memories shared across tools and sessions, so facts are retrieved on demand instead of being repeated.

Enables Claude to brief self-hosted GrokBot bots through their webhook routines and then collect their replies in the same thread, with tools for waiting on replies, checking the inbox, and reading conversations. It includes a self-hosted dashboard for registering bots, managing encrypted webhook keys and tokens, and connecting Claude apps over OAuth or API tokens.

Enables Claude to ingest WhatsApp chat exports and shared files, search messages without diacritics, read conversations and documents or images, and draft WhatsApp replies via wa.me links while keeping data in the user's own Cloudflare account.

Enables MCP-compatible AI agents to interact with Discord through a local Vencord plugin inside the native desktop client, providing over 30 tools for messaging, server management, roles, moderation, and invites without bot or user tokens in config files.

Enables starting, monitoring, stopping, and resuming detached agent runs with per-pool concurrency limits and run history.

Enables a dedicated WeChat account to route incoming messages as MCP events to ChatGPT Work automations, which can generate and send idempotent replies back to the original sender using signed webhooks and persisted pending queues.

Enables AI agents and LLM platforms to create and manage live 3D isometric sessions, upsert agents, update progress and status, trigger speech bubbles, and remove agents through stateful MCP tools over STDIO or HTTP.

Enables Claude to search a CRM account's contacts and deals, add notes, and create follow-up tasks, exposing read-only and additive-write tools over a bearer-token-protected Streamable HTTP endpoint. No tool can delete or edit existing records.

Enables MCP hosts that require a local stdio process to connect to the hosted public demo, forwarding tool discovery and calls for fictional scheduling scenarios, product information, draft-sharing preparation, and optional feedback.

Enables Claude Code users to inventory the MCP servers, skills, commands, agents and plugins installed on their machine, estimate how many tokens each one adds to every request, and switch them off per project or globally, including across other AI coding tools. It also monitors live sessions to report real token usage and the savings achieved from disabled items.