Self-Improving Development Rules with LLM Tools

A development workflow system that learns from mistakes, maintains task context, and provides powerful LLM and web interaction tools.

Overview

This skill enables AI assistants to maintain persistent knowledge across interactions, organize complex tasks, and leverage external tools (LLM APIs, web scrapers, search engines) to enhance development capabilities.

Core Workflow

1. Lessons Learned System

When you encounter reusable information or corrections:

Record library versions, model names, configuration details

Document fixes to mistakes or corrections received

Update the `Lessons` section in `.cursorrules` to prevent repeating errors

2. Scratchpad for Task Management

Use `.cursorrules` as a scratchpad to:

Review context at the start of each task

Clear outdated task information from previous work

Explain the current task clearly

Plan steps with todo markers:

```

[X] Completed task

[ ] Pending task

```

Update progress after each subtask completion

Reflect on milestones and plan next steps

Maintain both big picture and granular progress tracking

3. Tool Integration

All tools use `python3` from the virtual environment at `./venv`.

#### LLM API

Query language models for assistance:

```bash

venv/bin/python3 ./tools/llm_api.py --prompt "Your question" --provider "anthropic"

```

**Supported Providers:**

OpenAI (gpt-4o)

Azure OpenAI (configured via .env)

DeepSeek (deepseek-chat)

Anthropic (claude-3-sonnet-20240229)

Gemini (gemini-pro)

Local LLM (Qwen/Qwen2.5-32B-Instruct-AWQ)

**Programmatic Usage:**

```python

from llm_api import query_llm

response = query_llm("Your prompt", provider="anthropic")

```

#### Screenshot Verification

Capture and verify web pages visually:

```bash

Capture screenshot

venv/bin/python3 tools/screenshot_utils.py URL --output screenshot.png --width 1920 --height 1080

Verify with LLM

venv/bin/python3 tools/llm_api.py --prompt "Describe this page" --provider openai --image screenshot.png

```

**Programmatic Workflow:**

```python

from screenshot_utils import take_screenshot_sync

from llm_api import query_llm

screenshot_path = take_screenshot_sync('https://example.com', 'screenshot.png')

response = query_llm(

"What is the background color and title?",

provider="openai",

image_path=screenshot_path

)

```

#### Web Scraping

Extract content from multiple URLs concurrently:

```bash

venv/bin/python3 ./tools/web_scraper.py --max-concurrent 3 URL1 URL2 URL3

```

#### Search Engine

Search the web and get structured results:

```bash

venv/bin/python3 ./tools/search_engine.py "your search keywords"

```

**Output Format:**

```

URL: https://example.com

Title: Result title

Snippet: Result snippet

```

Combine with web scraper to fetch full page content.

Project-Specific Lessons

Python Environment

Always use the venv at `./venv`

Check for `uv` package manager: `which uv`

If available: activate venv, then `uv pip install <package>`

Otherwise: use standard `pip`

Debugging

Include debug information in program outputs

Always read files before editing them

Git/GitHub Workflows

For multiline commit messages, write to a file first

Use `git commit -F <filename>` to submit

Delete the temporary file afterward

Prefix commits/PRs with `[Cursor]`

Database Migrations (Laravel)

For new fields on existing tables, modify the original migration

Run `php artisan migrate:fresh --seed` instead of creating new migrations

Maintains clearer database structure

Character Encoding

Handle UTF-8 properly for international search queries

Send debug info to stderr, main output to stdout for pipeline integration

Matplotlib/Seaborn

Use `'seaborn-v0_8'` style name instead of `'seaborn'`

OpenAI Models

Use `'gpt-4o'` for GPT-4 with vision capabilities

Search Queries

For recent news, use current year (2025) or keyword "recent"

Step-by-Step Instructions

1. **Start New Task:**

- Open `.cursorrules` and review Scratchpad

- Clear outdated task context if necessary

- Write task explanation and initial plan with todo markers

2. **During Task Execution:**

- Update todo markers as subtasks complete

- Document any reusable discoveries in Lessons section

- Consult Scratchpad before planning next steps

3. **Using Tools:**

- Activate venv before running Python tools

- Use LLM API for quick queries or programmatically for complex workflows

- Use screenshot verification for visual testing

- Use web scraper and search engine for research tasks

4. **At Milestones:**

- Reflect on progress in Scratchpad

- Plan next phase based on big picture

- Update Lessons with any corrections or insights

5. **Task Completion:**

- Mark all todos as complete

- Document key learnings in Lessons

- Commit changes with proper message format

Notes

API keys may not be configured; if so, screenshot verification and LLM sections are disabled until configured

All Python tools support `--help` for detailed usage

Batch processing can be done by importing tool modules directly in Python scripts

Self-Improving Development Rules with LLM Tools

Self-Improving Development Rules with LLM Tools

Overview

Core Workflow

1. Lessons Learned System

2. Scratchpad for Task Management

3. Tool Integration

Capture screenshot

Verify with LLM

Project-Specific Lessons

Python Environment

Debugging

Git/GitHub Workflows

Database Migrations (Laravel)

Character Encoding

Matplotlib/Seaborn

OpenAI Models

Search Queries

Step-by-Step Instructions

Notes

Reviews (0)