AICreateAI — Data skill for Claude Code
AICreateAI: Claude Code skill for AI-to-AI data generation and distillation - zero-trust verified samples, diversity, adversarial review against model collapse.
How to install AICreateAI
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open omer-faruk-g/AICreateAI and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What AICreateAI does
AICreateAI: Claude Code skill for AI-to-AI data generation and distillation - zero-trust verified samples, diversity, adversarial review against model collapse. Prompt-only, one SKILL.md.
Alternatives in Data
- A Stock Data — A股全栈数据工具包 · 十一层架构 · 54端点 · 19数据源 · 零鉴权 Full-stack China A-share data toolkit for AI agents — 11 layers, 54 end 9.4k ★
- Context Mode — Benchmark Results — Benchmarked against real outputs from popular Claude Code MCP servers, Skills, and dev tools 5.6k ★
- MCP Server Chart — 🤖 A visualization mcp & skills contains 25+ visual charts using @antvis 4k ★
README
AICreateAI
AICreateAI is a skill for [Claude Code](https://code.claude.com) that governs AI-to-AI data generation, distillation and synthetic code. It is prompt-only: one `SKILL.md`, with no scripts and no hooks. It targets three failure modes that appear when one model produces training data for another: model collapse, the photocopy effect (errors copied into every later generation) and self-preference bias (a model rating its own output too highly).
Rules
| Section | Rule | What the model does |
|---|---|---|
| 1.1 | Zero-trust synthetic data | It never marks a generated sample as valid on its own confidence. Every code sample needs valid syntax with no compile errors or warnings, edge-case asserts or unit tests, and explicit error paths with no silent catch-all or empty blocks. |
| 1.1 | Reject unverified output | Code that cannot be compiled or checked by dry run or AST check is discarded. |
| 1.2 | Diversity enforcer | No repeated variable names, boilerplate or trivial templates across samples. At least 25 % of samples cover boundary conditions, race conditions, memory limits and non-standard inputs. |
| 2.1 | Adversarial self-review | Before a sample is final, it switches to an auditor role, tries to break the code with malicious and edge-case inputs, checks memory safety and time complexity, and rejects samples with hidden assumptions or unhandled cases. |
| 2.2 | Signal-to-noise filter | It strips prose, opinions and non-deterministic text and keeps only [Problem Specification] -> [Chain-of-Thought Reasoning Tree] -> [Verified Code] -> [Test Suite]. |
| 3 | Distillation standard | The student model gets dense, high-signal logic, not the teacher's verbose explanations. Reasoning always comes before the final code. |
| 4 | Pre-flight checklist | Before approving a sample, prompt or pipeline step it checks for errors, the adversarial review, leftover filler and missing tests. |
Install
git clone https://github.com/omer-faruk-g/AICreateAIRelated Skills
Capture Pair
the user (or a review pass) just corrected something a model produced. Preserve the before→after pair for the
Cross Model Agents
Bidirectional adversarial review between Claude Code (Opus 4.7) and Codex CLI (GPT-5.5). 31 agents, 4 skills,
Signum
Evidence-driven development pipeline with multi-model code review. Generates code against a contract, audits w
Dbt Model Planner
Claude Code skill: Plan and build dbt models from schema requirements through structured planning with human r
Agent Review Pipeline
Autonomous code review pipeline for Claude Code. Asymmetric dual-engine dispatch — Codex (correctness + advers
Salary Calculator
Runs the full Salary Calculator pipeline — research prompt generation, data validation, international adjustme
Related Agents
Security Boss
You need to identify vulnerabilities, audit auth flows, review code for security flaws, or threat-model a syst
Security Architecture
Boucle implements a defense-in-depth security model to protect against prompt injection attacks and maintain t
Prompt Forager
Context-starved prompt forager for diversity injection. Receives ONE line (a purpose) and deliberately nothing
