Project Coding Companion

Author: Ruben Granet Version: 1.0.0 License: MIT

An AI coding companion that coaches you to write better coding prompts before implementing — evaluate, clarify, plan, code, review. Instead of diving straight into code when a request is vague, it first helps you clarify what you actually need — then plans, implements, and reviews the changes.

Works with Claude Code, Claude.ai, Cursor, and OpenAI Codex.

What it does

• Coaches your prompts — evaluates every coding request against 5 dimensions (context, intent, scope, constraints, acceptance criteria) and asks targeted questions with explanations when something is missing.
• Respects conversation context — recognizes continuations ("ok do it", "next step") and doesn't re-ask questions you already answered.
• Understands repositories — inspects project structure, identifies the stack, and builds a mental model of the codebase.
• Plans implementations — proposes short, step-by-step plans (1–7 steps) and waits for your approval.
• Implements changes — reads before writing, makes minimal diffs, preserves existing style.
• Reviews and summarizes — lists every file changed, explains why, and suggests tests or follow-up work.
• Stays safe — never auto-commits, never runs destructive commands without confirmation.

Works with any language or stack: TypeScript, Python, Swift, Kotlin, Rust, Go, Java, Ruby, C#, and more.

When to use it

Great fit:

• You have an existing repository and want help adding features, fixing bugs, refactoring, writing tests, or understanding the code.
• You're learning to write better prompts for AI coding assistants.
• You want a disciplined workflow (evaluate prompt → coach → plan → implement → review) instead of ad-hoc code generation.

Not the best fit:

• Pure greenfield scaffolding with no existing code.
• Non-code tasks (documents, presentations, spreadsheets).
• Heavy frontend design work driven primarily by aesthetics.

Platform support

This repo ships three files — same instructions, adapted for each platform:

Folder structure

project-coding-companion-skill/        ← This repo
├── README.md                          ← You are here (for humans on GitHub)
├── LICENSE                            ← MIT
├── project-coding-companion/          ← Claude Skill folder
│   └── SKILL.md                       ← Claude Code & Claude.ai
├── .cursorrules                       ← Cursor
└── AGENTS.md                          ← OpenAI Codex

Note: The README.md lives at the repo root for GitHub readers. It is not inside the skill folder — per Anthropic's Skill specification, the skill folder contains only SKILL.md and optional scripts/, references/, assets/ directories.

Installation

Quick install

npx skills add rgranet/project-coding-companion-skill

Claude Code

1. Clone this repo:
   hljs language-bash

   Copy

   git clone https://github.com/rgranet/project-coding-companion-skill.git
   

2. Place the project-coding-companion/ folder in your Claude Code skills directory.
3. The skill activates automatically on coding requests.

Claude.ai (web & desktop)

1. Zip the skill folder (not the whole repo):
   hljs language-bash

   Copy

   zip -r project-coding-companion.zip project-coding-companion/
   

2. Go to Settings → Capabilities → Skills.
3. Click Upload skill and select the zip file.

Cursor

1. Copy .cursorrules from this repo into the root of your project:
   hljs language-bash

   Copy

   cp .cursorrules /path/to/your/project/
   

2. Open your project in Cursor — the rules apply automatically to all conversations.

Important: .cursorrules must be at the root of the project you're working on, not in a separate skills folder. Each project gets its own copy.

OpenAI Codex

1. Copy AGENTS.md from this repo into the root of your project:
   hljs language-bash

   Copy

   cp AGENTS.md /path/to/your/project/
   

2. Codex reads AGENTS.md automatically as agent instructions.

Important: Like .cursorrules, AGENTS.md must be at the root of the project you're working on.

Example workflows

1. Vague request → coaching → implementation

You: "Add search to my app"

1. Detects 3+ missing dimensions, enters coaching mode.
2. Asks 3 targeted questions (which screen? what's searchable? real-time or on submit?) with "→ Why this matters" explanations.
3. You answer. Restates the task as a checklist for confirmation.
4. Explores the codebase, plans steps, implements, summarizes changes.
5. Ends with a prompting tip: "Next time, mentioning the target screen saves a round-trip."

2. Clear request → straight to code

You: "Add a isFavorite: Bool property to the Thought entity, a toggle button in ThoughtDetailView, and a FavoritesListView filtered on isFavorite == true. Follow existing MVVM pattern."

1. Evaluates: all 5 dimensions present. "Clear and complete — I have everything I need."
2. Plans 4 steps, implements with minimal diffs, summarizes.

3. Explicit prompt coaching

You: "I want to refactor my services layer but I don't know how to describe it. Help me write a good prompt."

1. Asks you to describe the pain points in your own words.
2. Shows a 5-dimension scorecard (✅/❌/⚠️).
3. Proposes a rewritten prompt with all dimensions filled.
4. Explains key improvements so you can apply the same thinking next time.

Prompt examples

Add search to my app.

Help me understand the architecture of this project.

Fix the bug where deleting a parent entity doesn't cascade to children.

Help me write a better prompt for refactoring my networking layer.

Refactor the authentication middleware to use a strategy pattern instead
of the current if/else chain. Keep backward compatibility with existing
callers. Add unit tests for each strategy.

Configuration / Customization

The companion is language- and framework-agnostic. It discovers conventions from the project itself. You can customize it by:

• For Claude — edit SKILL.md directly, or add references/ and scripts/ directories inside the skill folder. Files in these directories are only loaded when relevant (progressive disclosure).
• For Cursor — edit .cursorrules at your project root. You can add project-specific rules (e.g., "always use Vitest instead of Jest", "all API routes go through the createHandler wrapper").
• For Codex — edit AGENTS.md at your project root. Same customization approach as Cursor.

Examples of things you can tune: max plan steps, coaching aggressiveness, testing expectations, preferred patterns, language-specific conventions.

Contributing

Contributions are welcome!

1. Open an issue to report bugs, suggest trigger phrases, or propose new behaviors.
2. Submit a PR — one concern per PR, with a clear description of what and why.
3. Test your changes by loading the companion in your platform of choice and running representative prompts.

When contributing, please update all three files (SKILL.md, .cursorrules, AGENTS.md) to keep them in sync.

License

MIT — see LICENSE.