DEV Community

Cover image for I Built a Free Toolkit That Makes Any AI Coding Agent Write Better Code
subhansh
subhansh

Posted on

I Built a Free Toolkit That Makes Any AI Coding Agent Write Better Code

#ai #agentskills #productivity #opensource

I Extracted 95+ Skills From Claude Fable 5, GPT-5.5, and Gemini CLI Into One Free Repo

Your AI coding agent is powerful but generic. It writes code with purple gradients, opens responses with "Great question!", and ships skeletons instead of working code.

I fixed that.

Agent Maxxing is a free, open-source collection of 95+ production-ready skills, 19 UI components, and 7 system prompts — extracted from the leaked system prompts of the world's best AI agents.

The Problem

Every AI coding agent has the same personality:

  • Starts every response with praise ("Great question!", "Absolutely!")
  • Writes code with no error handling ("Here's a skeleton, you fill it in")
  • Uses purple-to-blue gradients everywhere
  • Sounds like a corporate press release
  • Asks permission for low-risk actions instead of just doing them

The root cause? Their system prompts are generic. They don't know your design system, your coding standards, or how to sound like a human.

The Solution

I extracted the actual system prompts from:

  • Claude Fable 5 (3,826 lines) — personality, memory system, tone, refusal handling
  • GPT-5.5 Codex (11,104 lines) — engineering judgment, frontend rules, formatting
  • Gemini CLI (254 lines) — context efficiency, sub-agent orchestration
  • Claude Code (1,798 lines) — tool usage, code review methodology, agent delegation

Then I organized everything into 9 skill categories with 95+ individual skill files.

What's Inside

Engineering (22 skills)

  • Deep Code Review — 8-angle methodology from Claude Code bundled skills
  • Security Review — Senior security engineer audit with false-positive filtering
  • Execution Protocol — "Solve it, don't ask about it" problem-solving chain
  • Debugging Patterns — Systematic methodology for finding and fixing bugs
  • API Design — REST patterns, status codes, pagination
  • Database Patterns — Schema design, indexing, migrations
  • Performance Patterns — Lazy loading, caching, memoization

Frontend Design (16 skills)

  • Anti-Patterns — What makes AI output look AI-generated
  • Artifact Design — Deliberate design choices from Claude Code
  • Motion Design — Timing, easing, enter/exit patterns
  • Typography — Font pairing, scale, typographic systems
  • Responsive Design — Breakpoints, fluid layouts

UI Components (19 patterns)

  • Glass Card, Gradient Button, Modal Dialog, Tabs, Dropdown, Avatar, Progress Bar, Badge, Toggle Switch, Accordion, Tooltip, Navbar, Data Table, Loading Skeleton, Toast Notification, Breadcrumb, Pagination, Animated Input

System Prompts (7 modules)

Extracted and intelligently curated from leaked prompts:

  • Agent Core Personality — From Claude Fable 5
  • Coding Excellence — From GPT-5.5 Codex
  • Reasoning & Planning — From Codex Plan Mode
  • Frontend Mastery — From GPT-5.5 + Claude Design
  • Agent Orchestration — From Claude Code + Gemini CLI
  • Tone & Communication — From Fable 5 + Codex + Cursor

How It Works

Installation 

# Clone the repo
git clone https://github.com/subhansh-dev/agent-maxxing.git

# Install for your agent
cp -r agent-maxxing ~/.claude/skills/agent-maxxing  # Claude Code
cp -r agent-maxxing ~/.codex/skills/agent-maxxing   # Codex CLI
cp -r agent-maxxing .cursor/skills/agent-maxxing    # Cursor
Enter fullscreen mode Exit fullscreen mode

The Lazy Prompt

Paste this to your agent and it self-fine-tunes:

Deep dive into the agent-maxxing folder. Read every .md file — every skill, every system prompt, every component, every workflow example. Understand what each file teaches. Then fine-tune yourself: adopt the patterns, internalize the anti-patterns, apply the engineering judgment, use the writing style. Integrate all 95+ skills so you can use them on any task. From now on, before responding to anything, check if a relevant skill exists and apply it.
Enter fullscreen mode Exit fullscreen mode

How Agent Self-Fine-Tuning Works

It's not weight training — it's context injection. When an agent reads a skill file:

  1. The file content enters the agent's context window
  2. The agent internalizes the patterns for that session
  3. It applies those patterns to every subsequent response
  4. The behavior changes without retraining

The more skills it reads, the better it gets. The FINE-TUNE-AGENT.md walks the agent through every skill category in order.

Key Anti-Patterns (What Agents Stop Doing)

After fine-tuning, your agent will:

  • Stop opening with "Great question!" or "Absolutely!"
  • Stop using purple-to-blue gradients in every design
  • Stop shipping skeleton code ("Here's a basic implementation")
  • Stop hedging with "It seems like..." or "It appears that..."
  • Stop asking permission for low-risk actions
  • Start writing code with proper error handling
  • Start matching the existing codebase style
  • Start sounds like a human, not a press release
  • Start reviewing code with 8-angle methodology
  • Start shipping working code, not skeletons

The Golden Rules

Every top agent in this collection follows these:

  1. Search before answering — Never guess when you can check
  2. Read the codebase first — Don't assume, investigate
  3. Parallel tool calls — Independent operations run simultaneously
  4. Match existing patterns — Follow the repo's conventions
  5. Show, don't tell — Demonstrate with examples
  6. Be concise — Short paragraphs, flat lists, no filler
  7. Take ownership — Fix mistakes, don't deflect
  8. Respect the user's time — Don't ask what you can figure out
  9. Ship working code — Not skeletons, not "you'll need to add..."
  10. Sound human — Not like a press release or corporate blog

Compatible With 60+ Agents

Works with any agent supporting the Agent Skills specification:

  • Claude Code
  • Codex CLI
  • OpenCode
  • Cursor
  • Continue
  • Kilo Code
  • OpenClaw
  • Pi Agent
  • Hermes
  • Gemini CLI
  • And 50+ more

What Makes This Different

Agent Maxxing Generic Skill Packs
Source Extracted from Claude Fable 5, GPT-5.5, Gemini CLI Community-written prompts
Depth 95+ skills covering engineering, design, writing, security 5-10 surface-level tips
Components 19 production-ready UI patterns with CSS None
System Prompts 7 fine-tuned personality modules None
Anti-Slop Dedicated anti-patterns + writing style guide "Be helpful"

Try It Now

  1. Clone the repo
  2. Install for your agent
  3. Paste the lazy prompt
  4. Watch your agent transform 
github.com/subhansh-dev/agent-maxxing
Enter fullscreen mode Exit fullscreen mode

MIT licensed. Zero dependencies. Just markdown.

Built by a 17-year-old who got tired of generic AI output.

Top comments (0)