Cursor AI Best Practices for podcast_scraper¶

🚀 Quick Reference: Model Selection¶

Default Rule of Thumb:

Auto → fast iteration, small/local changes
Manual model selection → planning, debugging, architecture, CI failures

Auto optimizes for speed & cost, not deep reasoning.

🎯 Quick Reference: @Rules System¶

Two separate rule systems:

System	How Loaded	When to Use
`.cursorrules`	Automatic (always active)	Core rules for every interaction
`.cursor/rules/*.mdc`	Manual via `@Rules filename`	Specialized context when needed

Load @Rules for specific tasks:

@Rules ai-guidelines        # Detailed commit/push checklists
@Rules markdown-style       # Markdown formatting guide
@Rules testing-strategy     # Testing patterns and pytest markers
@Rules git-worktree         # Worktree workflow commands
@Rules module-boundaries    # Architecture constraints

Example:

# Starting to write tests

@Rules testing-strategy
Write unit tests for the new provider

# Creating worktree

@Rules git-worktree
Set up worktree for issue #200

Model Roles (Quick Guide)¶

Fast / Mechanical Tasks (small, obvious changes):

Rename functions, generate boilerplate/tests, minor refactors, formatting & docstrings
Models: GPT-5.1 Codex Mini, Sonnet 3.5, Gemini 3 Flash
@Rules: Usually none needed (.cursorrules handles it)

Main Coding & Reviews (daily driver):

Multi-file changes, writing tests, medium refactors, code reviews
Models: GPT-5.1 Codex Max (default), Sonnet 4.5
@Rules: Load as needed for context (@Rules testing-strategy, etc.)

Deep Reasoning / Hard Problems (correctness matters):

CI failures you can't reproduce, packaging/Docker/deps, architecture decisions, concurrency/async, edge cases, "Why is this happening?"
Models: GPT-5.2, Opus 4.5
@Rules: Always load relevant context (@Rules module-boundaries, etc.)

High-Risk / Strict Correctness (use sparingly):

Security-sensitive code, release prep, complex refactors
Models: GPT-5.1 Codex Max High
@Rules: @Rules ai-guidelines for critical workflows

How to Influence Cursor Auto¶

Control Scope (most important):
Small selection → fast model
Multiple files / logs → stronger model
CI logs + diff → Auto upgrades
Use Reasoning Language:
Words that push Auto upward: analyze, root cause, tradeoffs, why, alternatives
Words that keep it cheap: rewrite, generate, convert
Ask for Options:
"Give 2–3 possible fixes, explain tradeoffs, recommend one."
Split Planning and Execution:
Chat (no code selected): "Analyze and propose a plan. No code yet."
Inline edit (code selected): "Implement step 2 from the plan."
Use Composer for Big Tasks:
Composer sessions bias Auto toward stronger models and longer reasoning.
Load @Rules for Context:
@Rules testing-strategy for test-related tasks
@Rules module-boundaries for architecture work
@Rules git-worktree for worktree management

One-Line Mental Model¶

Auto for hands, manual for brain. Load @Rules for specialized context.

Current Situation¶

See also:

Development Guide - Implementation instructions

Testing Guide - Test execution commands

CI/CD - CI/CD pipeline details

Git Worktree Guide - Worktree workflow

Markdown Linting Guide - Markdown standards

This document captures best practices for using Cursor effectively in the podcast_scraper project, which is a Python-heavy codebase with:

AI Rules System: .cursorrules (always active) + .cursor/rules/*.mdc (on-demand via @Rules)
CI/CD: GitHub Actions workflows, Docker builds, Snyk security scans
Testing: Unit, integration, E2E tests with two-tier strategy (fast/slow)
Documentation: RFCs, PRDs, MkDocs site, markdown linting
Git Workflow: Worktree-based parallel development, never push to main
Strict guidelines: Mandatory commit approval, make ci-fast before commit

Problem: Without proper model selection and @Rules usage, Cursor can:

Choose suboptimal models for complex tasks (CI debugging, architecture decisions)
Miss project-specific patterns (mandatory commit approval, module boundaries)
Generate code that doesn't follow project standards (testing requirements, code style)
Waste time on mechanical tasks that could be faster
Lack specialized context for complex workflows (testing, worktrees, architecture)

Solution: Layered context system:

.cursorrules (287 lines) - Always active, enforces core rules
.cursor/rules/*.mdc (200-400 lines each) - Load with @Rules for specialized tasks
Model selection - Choose appropriate model for task complexity
Prompt templates - .cursor/prompts/ for structured prompts

Mental Model¶

Cursor is best thought of as:

Auto → speed-first, cost-aware assistant (good for mechanical work)
Manual model selection → correctness, reasoning, and design (required for complex tasks)
@Rules system → contextual knowledge loader (activate specialized guides when needed)

Key Principle: Auto optimizes for fast iteration. You must explicitly ask for depth and load specialized context with @Rules.

Recommended Workflows with @Rules¶

Workflow 1: Implementing New Feature (RFC-Based)¶

Scenario: Implementing RFC-025 metrics collection

Step 1: Design phase
Chat: "@Rules ai-guidelines
       Read RFC-025 and analyze implementation approach"

Step 2: Architecture check
Chat: "@Rules module-boundaries
       Review proposed changes for module boundary violations"

Step 3: Implementation
Composer: [Paste design-rfc.md template]
          "Implement Phase 1: Basic Metrics Collection"

Step 4: Testing
Inline Chat: "@Rules testing-strategy
             Write unit and integration tests for metrics"

Step 5: Documentation
Inline Chat: "@Rules markdown-style
             Update DEVELOPMENT_GUIDE.md with metrics usage"

Workflow 2: Debugging CI Failure¶

Scenario: make ci failure

Step 1: Root cause analysis
Chat: [Paste CI logs + git diff]
      "@Rules testing-strategy
       Analyze root cause of this CI failure"

Step 2: Fix implementation
Inline Chat: [Select failing test]
            "Fix based on analysis, following project patterns"

Step 3: Verify commit
Terminal: make ci-fast

Step 4: Commit
Chat: "@Rules ai-guidelines
       Show git status and git diff, wait for approval"

Workflow 3: Creating New Worktree¶

Scenario: Start work on issue #200

Step 1: Setup worktree
Chat: "@Rules git-worktree
       Set up worktree for issue #200 (add Dependabot)"

Step 2: Verify setup
Terminal: cd ../podcast_scraper-200-dependabot
          source .venv/bin/activate
          cursor .

Step 3: Daily work
[Work in isolated worktree with clean AI context]

Step 4: Cleanup after merge
Chat: "@Rules git-worktree
       Remove worktree and cleanup references"

Workflow 4: Refactoring Module¶

Scenario: Refactoring workflow pipeline

Step 1: Architecture review
Chat: "@Rules module-boundaries
       Analyze current workflow.py for boundary violations"

Step 2: Design refactor
Chat: [Paste design-rfc.md template]
      "Propose refactoring plan maintaining module boundaries"

Step 3: Implement
Composer: "@Rules module-boundaries
          Implement refactoring following the plan"

Step 4: Test coverage
Inline Chat: "@Rules testing-strategy
             Ensure test coverage for refactored code"

Workflow 5: Writing Documentation¶

Scenario: Update README and guides

Step 1: Content changes
Inline Chat: [Select markdown file]
            "Update with new feature information"

Step 2: Formatting
Chat: "@Rules markdown-style
       Review markdown for style issues"

Step 3: Validation
Terminal: make fix-md
          make lint-markdown
          make docs

Step 4: Commit
Chat: "@Rules ai-guidelines
       Show changes and wait for approval"

Pro Tips for @Rules Usage¶

1. Multiple @Rules at once

You can load multiple rules in one prompt:

@Rules testing-strategy @Rules module-boundaries
Write tests for the new module following architecture constraints

2. Rules persist in conversation

Once loaded, @Rules stay active in that conversation:

First message:
@Rules testing-strategy
Help me write tests

Later messages:
Add more test cases
[testing-strategy.mdc still loaded]

3. New chat = reset

Start new chat → Must load @Rules again
Previous @Rules don't carry over

4. Start without @Rules

For simple tasks, .cursorrules is usually enough:

You: "Fix this typo in config.py"
[.cursorrules handles this fine]

You: "@Rules testing-strategy
     Write comprehensive test suite"
[Now you need the detailed testing guide]

Recommended Workflow (Python Project)¶

GPT-5.2 → decide approach & edge cases
Codex Max → implement + tests
Sonnet 4.5 → review diff
Codex Max / High → apply fixes
Composer → PR description & checklist

Summary: One-Line Mental Model¶

Auto for hands, manual for brain.

Model Selection Strategy¶

Fast / Mechanical Work¶

Use when: Changes are local and obvious.

Examples for this project:

Renaming functions/variables
Generating boilerplate tests
Small refactors (single file)
Formatting or docstrings
Updating imports
Fixing simple linting errors

Typical models:

GPT-5.1 Codex Mini
Sonnet 3.5
Gemini 3 Flash

Project-specific use cases:

Running make format and applying fixes
Adding type hints to simple functions
Updating docstrings to match project standards
Generating test fixtures

Main Coding & Reviews (Daily Driver)¶

Use for: Most feature work in this project.

Examples for this project:

Multi-file changes (workflow, transcription, summarization modules)
Writing or updating tests (unit, integration, E2E)
Medium refactors (module reorganization)
Primary code reviews
Implementing RFC features
Adding new provider implementations

Typical models:

GPT-5.1 Codex Max
Sonnet 4.5

Project-specific use cases:

Implementing new transcription providers
Adding new summarization backends
Creating new test categories (acceptance tests)
Refactoring workflow pipeline
Implementing RFC-019 E2E test improvements
Adding metrics collection (RFC-025)

Deep Reasoning / Hard Problems¶

Use when: Understanding and correctness matter.

Examples for this project:

CI failures with unclear cause (GitHub Actions, Docker builds)
Docker / dependency issues (ML dependencies, version conflicts)
Architecture decisions (module boundaries, protocol design)
Concurrency or async bugs (multi-threaded downloads, parallel processing)
"Why is this happening?" (Whisper progress bar, test failures)
Complex refactors (pipeline refactoring, protocol extensions)

Typical models:

GPT-5.2
Opus 4.5

Project-specific use cases:

Debugging make ci failures
Analyzing test timeout issues
Understanding parallel test execution problems
Designing new protocol extensions
Resolving merge conflicts in complex refactors
Investigating Whisper transcription issues
Debugging Docker build failures

High-Risk / Strict Correctness¶

Use sparingly for sensitive changes.

Examples for this project:

Security-related code (secrets handling, API keys)
Release prep (version bumps, changelog)
Complex refactors (protocol changes, breaking API changes)
Critical bug fixes (data loss, corruption)

Typical models:

GPT-5.1 Codex Max High

Project-specific use cases:

Modifying .ai-coding-guidelines.md (affects all AI workflows)
Changing git workflow rules
Security audit fixes
Release preparation
Breaking API changes

Cursor Auto: How It Works (Practically)¶

Auto behavior:

Chooses model based on scope, context size, and task type
Prefers cheaper/faster models
Rarely escalates to top reasoning models unless forced

Auto tends to pick:

Small selection → fast model
Medium changes → Codex Max / Sonnet
Large context → Codex Max
Text-only → Gemini Flash

Auto almost never chooses GPT-5.2 or Opus 4.5 on its own.

How to Influence Auto (What Actually Works)¶

1. Control Scope (Most Important)¶

For this project:

Select one function → fast model (good for simple fixes)
Select multiple files → stronger model (good for multi-file refactors)
Paste CI logs + diff → Auto upgrades (essential for debugging make ci failures)
Select entire test file → stronger model (good for test refactoring)

Example:

❌ Bad: "Fix the CI failure" (no context)
✅ Good: Select CI log output + paste into chat + "Analyze root cause of this CI failure"

Words that push Auto upward:

analyze
root cause
tradeoffs
why
alternatives
risks
architecture
design decision

Words that keep it cheap:

rewrite
generate
convert
format
fix linting

Project-specific examples:

❌ "Fix the test failure"
✅ "Analyze root cause of this test timeout and propose solutions with tradeoffs"

❌ "Update the workflow"
✅ "Review this workflow change for architecture implications and potential risks"

"Give 2–3 viable approaches for implementing RFC-025 metrics collection, explain tradeoffs, recommend one."

Project-specific:

"Analyze this CI failure and propose 2-3 solutions. Consider: test execution time, parallel vs sequential, and CI/CD impact."

4. Split Planning and Execution¶

Recommended workflow for this project:

Chat (no code selected)
"Analyze RFC-023 and propose implementation plan. Do not write code yet."
"Review this GitHub issue and create a plan. Consider project guidelines."
Composer (for RFCs/PRDs)
Paste RFC template
Add context (feature, related RFCs)
Let Composer reason
Follow up with "Now implement step X"
Inline chat (code selected)
"Implement step 2 of the plan"
"Apply the recommended solution"

Project-specific use cases:

RFC creation: Use Composer with RFC template
Issue analysis: Use Chat to analyze and plan
Implementation: Use Inline Chat with selected code
Code review: Use Chat with diff selected

5. Use Composer for Big Tasks¶

Composer sessions:

Bias Auto toward stronger models
Encourage structured reasoning
Ideal for RFCs and multi-step features

Project-specific use cases:

Creating new RFCs (RFC-024, RFC-025, RFC-026)
Implementing complex features (E2E test improvements)
Multi-file refactors (pipeline refactoring)
Documentation updates (DEVELOPMENT_GUIDE.md expansion)

Understanding Cursor's Rule Systems¶

Two Separate Rule Systems¶

Cursor has two different systems for loading context:

1. `.cursorrules` (Root File) - AUTOMATIC ✅¶

Location: /.cursorrules (project root)

How it works:

✅ Loaded automatically at session start
✅ Always active for every AI interaction
✅ You do nothing - Cursor reads it automatically
✅ Priority: Highest - always in context

What it contains (287 lines):

Core git workflow rules (never push to main)
Testing strategy (make ci-fast before commit)
Code quality standards (imports, type hints, docstrings)
Module boundaries (architecture constraints)
Documentation requirements

Think of it as: Your project's "constitution" - always in effect.

2. `.cursor/rules/*.mdc` Files - MANUAL ON-DEMAND 📋¶

Location: /.cursor/rules/ folder

How it works:

❌ NOT loaded automatically - Cursor ignores them by default
✅ Loaded manually when you mention them with @Rules
✅ You control when - Load specific context when needed
✅ Scope: Current conversation only

How to load:

In Cursor Chat, type @Rules followed by filename (without .mdc):

@Rules ai-guidelines        # Loads detailed commit/push checklists
@Rules markdown-style       # Loads markdown formatting guide
@Rules testing-strategy     # Loads testing guide with examples
@Rules git-worktree         # Loads worktree workflow
@Rules module-boundaries    # Loads architecture constraints

What happens:

You type: @Rules testing-strategy
         ↓
Cursor loads: .cursor/rules/testing-strategy.mdc
         ↓
AI now has: .cursorrules (always) + testing-strategy.mdc (just added)

Available .mdc files:

ai-guidelines.mdc (289 lines) - Detailed commit/push checklists
markdown-style.mdc (128 lines) - Markdown formatting with examples
testing-strategy.mdc (254 lines) - Testing guide with pytest markers
git-worktree.mdc (309 lines) - Worktree workflow and commands
module-boundaries.mdc (238 lines) - Architecture constraints

Think of these as: "Power-ups" you activate for specific tasks.

When to Use Each¶

Use `.cursorrules` (Always Active)¶

✅ Rules that apply to every interaction ✅ Critical workflows (git, commits, pushes) ✅ Code standards that never change ✅ Architecture rules that must always be enforced

Use `@Rules *.mdc` (On-Demand)¶

✅ Specific workflows you're about to perform ✅ Detailed guides for complex tasks ✅ Context that's only sometimes relevant

Examples:

# About to write tests

@Rules testing-strategy
Write unit tests for the new provider

# About to create worktree

@Rules git-worktree
Set up worktree for issue #200

# About to edit documentation

@Rules markdown-style
Update the README with new features

# About to refactor modules

@Rules module-boundaries
Refactor the workflow pipeline

Comparison Table¶

Feature	`.cursorrules`	`.cursor/rules/*.mdc`
Who loads it?	Cursor (automatic)	You (manual via `@Rules`)
When?	Every session start	Only when you mention it
Scope	All interactions	Current conversation only
Control	Always on	You decide when to use
Size	287 lines (optimized)	200-400 lines each (detailed)
Purpose	Core project rules	Specialized workflows

Real Examples¶

Example 1: Normal Coding (No @Rules)¶

You: "Add a new function to rss/downloader.py"

AI has access to:
✅ .cursorrules (automatic)
❌ testing-strategy.mdc (not loaded)
❌ git-worktree.mdc (not loaded)

AI uses .cursorrules to:
- Follow module boundaries
- Use correct import order
- Add type hints

Example 2: Writing Tests (With @Rules)¶

You: "@Rules testing-strategy
     Help me write tests for the new downloader function"

AI has access to:
✅ .cursorrules (automatic)
✅ testing-strategy.mdc (you just loaded it)

AI uses BOTH to:
- Follow .cursorrules (always)
- Use testing-strategy.mdc for pytest markers, mock patterns, etc.

Example 3: Creating Worktree (With @Rules)¶

You: "@Rules git-worktree
     I need to create a new worktree for issue #200"

AI has access to:
✅ .cursorrules (automatic)
✅ git-worktree.mdc (you just loaded it)

AI provides:
- Specific worktree commands
- Branch naming with issue number
- Complete setup workflow

Why Have Both Systems?¶

Problem if everything was in .cursorrules:

File would be 2,000+ lines
Loads into EVERY conversation
Wastes context on irrelevant info
Slower session starts

Solution with layered approach:

Core rules always active (287 lines)
Specialized rules on-demand (load when needed)
Efficient context usage

Current Project Structure¶

.cursorrules                    # Always loaded (287 lines)
.cursor/
├── rules/                      # Load manually with @Rules
│   ├── ai-guidelines.mdc       # Commit/push checklists (289 lines)
│   ├── markdown-style.mdc      # Markdown guide (128 lines)
│   ├── testing-strategy.mdc    # Testing guide (254 lines)
│   ├── git-worktree.mdc        # Worktree workflow (309 lines)
│   └── module-boundaries.mdc   # Architecture (238 lines)
└── prompts/                    # Copy/paste templates
    ├── debug-ci.txt
    ├── design-rfc.md
    ├── code-review.txt
    └── implementation-plan.txt

Prompt Files in `.cursor/prompts/`¶

Manual Templates (Copy/Paste)¶

Prompt files are different from rules files:

Rules (.cursorrules and .mdc) = Context that Cursor loads
Prompts (.txt and .md) = Templates you copy/paste

Current structure:

.cursor/prompts/
├── debug-ci.txt          ✅ Created (Issue #95)
├── design-rfc.md          ✅ Created (Issue #95)
├── code-review.txt       ✅ Created (Issue #95)
└── implementation-plan.txt ✅ Created (Issue #95)

How to use:

Open the prompt file in .cursor/prompts/
Copy the entire content
Paste into Cursor Chat, Composer, or Inline Chat
Add your specific context (CI logs, code diff, etc.)
Send the prompt

These are NOT loaded by Cursor - you manually copy/paste them.

Available Prompt Templates¶

The following prompt templates are available in .cursor/prompts/:

`debug-ci.txt` - CI Failure Debugging¶

Purpose: Step-by-step analysis of CI failures with root cause identification.

Content: Analyzes CI failures systematically:

Summarizes what failed (job, step, error)
Lists 2-3 plausible root causes, ranked by likelihood
Explains how to confirm or rule out each cause
Recommends safest fix with minimal side effects
Calls out hidden risks and follow-up checks

Use when: CI fails, logs are long, cause unclear Effect: Auto almost always upgrades to a reasoning-heavy model (GPT-5.2/Opus 4.5)

Location: .cursor/prompts/debug-ci.txt

`design-rfc.md` - RFC and Feature Design¶

Purpose: Structured approach to designing RFCs, features, and refactors.

Content: Guides through design process:

Restates problem in own words
Defines explicit goals and non-goals
Proposes 2-3 viable designs with tradeoffs (complexity, risk, testability)
Recommends one design with justification
Outlines concrete implementation plan
Identifies risks and mitigation strategies
Proposes test strategy

Use when: Starting features, refactors, or PRDs Effect: Pushes Auto to Opus / GPT-5.2-level reasoning

Location: .cursor/prompts/design-rfc.md

`code-review.txt` - Code Review¶

Purpose: Systematic code review focusing on correctness, tests, and maintainability.

Content: Reviews code or diffs with focus on:

Correctness and edge cases
Missing or weak tests
API and behavior consistency
Readability and maintainability

Output: Bullet list of findings labeled as MUST / SHOULD / NICE-TO-HAVE

Use when: Reviewing your own diff or as a second reviewer Effect: Keeps reviews sharp without overengineering

Location: .cursor/prompts/code-review.txt

`implementation-plan.txt` - Implementation Planning¶

Purpose: Create structured implementation plans before writing code.

Content: Checklist-style planning:

Lists files/modules that will change
Describes changes per file
Identifies edge cases and failure modes
Specifies tests to add or update
Notes CI, Docker, or packaging implications

Use when: Changes span multiple files or systems Effect: Auto upgrades model + prevents premature coding

Location: .cursor/prompts/implementation-plan.txt

How to Use Prompt Templates¶

Open the prompt file in .cursor/prompts/
Copy the entire content
Paste into Cursor Chat, Composer, or Inline Chat
Add your specific context (CI logs, code diff, feature description, etc.)
Send the prompt - Cursor will use appropriate model based on prompt structure

2. **Codex Max (Composer)**
   - Paste `design-rfc.md` template
   - "Implement Phase 1: Basic Metrics Collection from JUnit XML"

3. **Sonnet 4.5 (Inline Chat)**
   - Select generated code
   - "Review this implementation for correctness and project patterns"

4. **Codex Max (Inline Chat)**
   - "Apply review feedback and ensure tests pass"

5. **Composer**
   - "Generate PR description with checklist based on RFC-025 Phase 1"

Subagents and Commands¶

What Are Subagents?¶

Subagents are separate AI agents the main Cursor agent can delegate to. Each has its own context; long or noisy work (e.g. full CI, acceptance tests) runs there and the main chat gets a summary. Rule 9 (no background make in main chat) applies to the main agent; a subagent running make ci in its own context returns a short result and does not violate that. All subagents use the project venv only: run from project root and use Makefile (which uses .venv/bin/python); never use global python/pip/pytest/npx.

Aspect	Meaning
Own context	Subagent's logs and output stay in its context; main chat stays clean.
Foreground / background	Foreground = main agent waits for result. Background = you can keep working while subagent runs.
Built-in	Explore (codebase search), Bash (shell commands), Browser (MCP). Cursor uses these when appropriate.
Custom	`.cursor/agents/*.md` (project) or `~/.cursor/agents/` (user) with YAML frontmatter and prompt.

Commands vs Subagents¶

	Commands (`/name`)	Subagents
What	You type `/verify` etc.; Cursor injects the command's Markdown as the prompt. Main agent does the work in the same chat.	Main agent delegates to another agent; that agent has its own context and returns a summary.
Where	`.cursor/commands/*.md` or `~/.cursor/commands/`	Built-in (Explore, Bash, Browser) or custom `.cursor/agents/*.md`
Best for	Repeatable start of a workflow (verify, review, pr, debug-ci, rfc) in one conversation.	Long/noisy runs (full CI, acceptance tests), deep search, parallel work, dedicated verifier.

Use a command when you want the main agent to do the work in this chat (e.g. /review, /pr).
Use a subagent when you want context isolation or a short summary (e.g. "run ci and tell me the result," "run acceptance tests for planet money").

Project Custom Subagents (`.cursor/agents/`)¶

This project defines three custom subagents. All run from project root and use the project venv only (Makefile uses .venv when present).

Subagent	File	When to use
Verifier	`verifier.md`	"Verify before commit," "run ci and report," "is the tree green?" Runs format-check, lint, lint-markdown, full make ci; optionally docker-test. Returns PASSED/FAILED + short summary.
CI Fix Loop	`ci-fix-loop.md`	"Run ci and fix until it passes," "make ci green." Runs full make ci; on failure fixes and re-runs up to 3 times. Returns final status + what was fixed.
Acceptance	`acceptance.md`	"Run acceptance tests," "run all acceptance configs." Runs `make test-acceptance CONFIGS="config/acceptance/*.yaml"` (or user pattern). Returns Status, Session ID, Summary; suggests `make analyze-acceptance SESSION_ID=…`.

Optional later: Docs check (lint-markdown + docs), PR prep (status/diff + docker-test). Use Cursor's built-in Explore for "find all X in codebase."

Subagents vs Skills¶

Subagents = separate context and/or parallel work (verify, run CI and fix, run acceptance).
Skills = procedures the main agent follows in the same context (commit-with-approval, push-to-pr, efficient pytest). Use skills for single-shot, repeatable procedures; use subagents when you want isolation or summarization.

Project-Specific Recommendations¶

1. Always Use @Rules for Specialized Tasks¶

Load appropriate context when needed:

Task	@Rules to Load	Why
Writing tests	`@Rules testing-strategy`	Pytest markers, mock patterns, coverage
Creating worktree	`@Rules git-worktree`	Setup commands, branch naming, cleanup
Refactoring modules	`@Rules module-boundaries`	Architecture constraints, SRP
Editing markdown	`@Rules markdown-style`	Formatting rules, auto-fix commands
Committing changes	`@Rules ai-guidelines`	Detailed commit/push checklists

Example:

❌ Bad: "Write tests for rss/downloader.py"
✅ Good: "@Rules testing-strategy
        Write tests for rss/downloader.py with appropriate pytest markers"

❌ Bad: "Update README.md"
✅ Good: "@Rules markdown-style
        Update README.md following project markdown standards"

2. Core Rules Are Always Active¶

You don't need to mention these - .cursorrules handles them:

✅ Git workflow (never push to main) ✅ Basic code standards (imports, type hints) ✅ Testing requirement (make ci-fast before commit) ✅ Module boundaries (high-level) ✅ Documentation requirements

Example:

You: "Add a new function to config.py"

AI automatically:
- Follows module boundaries (.cursorrules)
- Adds type hints (.cursorrules)
- Uses correct import order (.cursorrules)
[No @Rules needed for basic standards]

3. Mandatory Workflow Steps (Enforced by .cursorrules)¶

Before committing:

Run make ci-fast (pre-commit hook will also run this)
Show git status (mandatory)
Show git diff (mandatory)
Wait for explicit approval (mandatory)
Get commit message from user (mandatory)

Before pushing to PR:

Pre-commit hook passed during commit
Show git status (mandatory)
Show git diff or summary (mandatory)
Conditionally run make docker-test (if Docker-related changes)
Wait for explicit approval (mandatory)

Example prompt:

@Rules ai-guidelines
Commit these changes following the mandatory workflow

4. Test Execution Strategy (Use @Rules testing-strategy)¶

For this project:

# Default: Fast tests first

make ci-fast  # ~6-10 min (unit + fast integration + fast e2e)

# If ML code changed: Add slow tests

make test-integration-slow  # Whisper, summarization, speaker detection
make test-e2e-slow         # Full E2E with ML models

# Before final PR: Full validation

make ci  # ~10-15 min (includes coverage)

I changed providers/ml/whisper_utils.py. What tests should I run?

AI response:

1. make ci-fast (always)
2. make test-integration-slow (because ML code changed)
3. Verify specific test: pytest tests/integration/test_whisper_utils.py -v

5. Documentation Updates (Use @Rules markdown-style)¶

When implementing features:

Update relevant RFCs (Draft → Accepted → Completed)
Update guides if adding new patterns
Update mkdocs.yml if adding new docs
Run make fix-md before committing

Example with @Rules:

@Rules markdown-style
Update README.md with new provider feature and run markdown validation

6. Git Worktree Workflow (Use @Rules git-worktree)¶

For parallel development:

@Rules git-worktree
Create worktree for issue #200 with proper branch naming

AI provides:
- make wt-setup command
- Branch naming: feat/200-description
- Isolated venv setup
- Cursor instance management

One branch = One worktree = One Cursor window
Clean AI context per task
No branch switching
Parallel development

7. Model Selection for Common Tasks¶

With @Rules integration:

Task	Model	@Rules
RFC/PRD creation	GPT-5.2, Opus 4.5	`@Rules module-boundaries`
Test implementation	Codex Max	`@Rules testing-strategy`
CI debugging	GPT-5.2	`@Rules testing-strategy`
Module refactoring	GPT-5.2	`@Rules module-boundaries`
Documentation	Codex Max	`@Rules markdown-style`
Worktree setup	Auto or Codex Max	`@Rules git-worktree`
Commit/push	Auto	`@Rules ai-guidelines`

8. Always Reference Project Guidelines¶

Core documentation to mention:

# For architecture decisions

@Rules module-boundaries
Consider ARCHITECTURE.md for this refactoring

# For testing

@Rules testing-strategy
Follow TESTING_GUIDE.md patterns

# For CI/CD

Check docs/ci/index.md for CI pipeline details

# For markdown

@Rules markdown-style
Follow MARKDOWN_LINTING_GUIDE.md standards

Making This Fast (Recommended)¶

Option A: Split View¶

Left pane: .cursor/prompts/
Right pane: Chat / Composer

Copy → paste becomes muscle memory.

Option B: VS Code / Cursor Snippets (Best UX)¶

Convert prompts into snippets:

Type ci-debug → Enter (pastes debug-ci.txt)
Type rfc-design → Enter (pastes design-rfc.md)
Type impl-plan → Enter (pastes implementation-plan.txt)
Type code-review → Enter (pastes code-review.txt)

This removes copy/paste entirely.

Option C: Quick Access to Guidelines¶

Create snippet for project guidelines:

Type guidelines → Shows key rules from .ai-coding-guidelines.md
Type test-strategy → Shows test requirements
Type commit-workflow → Shows mandatory commit steps

Key Takeaways¶

General Principles¶

Auto for hands, manual for brain
Auto is good, but conservative
Prompt structure matters more than model picker
Plan first, code second
Use multiple models intentionally, not randomly

@Rules System Principles¶

.cursorrules is always active - Core project rules loaded automatically
.mdc files are on-demand - Load with @Rules filename when needed
Start without @Rules - .cursorrules handles most cases
Add @Rules for specialized tasks - Testing, worktrees, markdown, architecture
Multiple @Rules allowed - @Rules testing-strategy @Rules module-boundaries
New chat = reset - Must reload @Rules in new conversations

Project-Specific Principles¶

Always reference .cursorrules - Automatically enforces critical rules
Use @Rules for specialized context - Load detailed guides when needed
Never commit without showing changes and getting approval - Enforced by .cursorrules
Always run make ci-fast before committing - Pre-commit hook also runs this
Use appropriate test execution mode - Fast tests → slow tests (if ML changed) → full CI
Update documentation when implementing features
Follow RFC/PRD process for new features
Use worktrees for parallel development - @Rules git-worktree for setup

@Rules Quick Reference¶

When You're...	Load This	Why
Writing tests	`@Rules testing-strategy`	Pytest markers, mock patterns
Creating worktree	`@Rules git-worktree`	Setup commands, branch naming
Refactoring modules	`@Rules module-boundaries`	Architecture constraints
Editing markdown	`@Rules markdown-style`	Formatting, auto-fix commands
Committing changes	`@Rules ai-guidelines`	Detailed commit checklist
Regular coding	(none needed)	.cursorrules handles it

One-Line Rule¶

Auto for hands, manual for brain. For this project: .cursorrules always on, @Rules when specialized.

Next Steps¶

✅ Create prompt templates in .cursor/prompts/ (Completed - Issue #95):
✅ debug-ci.txt
✅ design-rfc.md
✅ code-review.txt
✅ implementation-plan.txt
Set up snippets in Cursor/VS Code for quick access:
Convert prompts into VS Code snippets for instant access
Type ci-debug → Enter (pastes debug-ci.txt)
Type rfc-design → Enter (pastes design-rfc.md)
Type impl-plan → Enter (pastes implementation-plan.txt)
Type code-review → Enter (pastes code-review.txt)
Practice workflows with common tasks:
RFC implementation (use design-rfc.md)
CI debugging (use debug-ci.txt)
Code reviews (use code-review.txt)
Multi-file changes (use implementation-plan.txt)
Review and update this guide based on experience

Cursor AI Best Practices for podcast_scraper¶

🚀 Quick Reference: Model Selection¶

🎯 Quick Reference: @Rules System¶

Model Roles (Quick Guide)¶

How to Influence Cursor Auto¶

One-Line Mental Model¶

Current Situation¶

Mental Model¶

Recommended Workflows with @Rules¶

Workflow 1: Implementing New Feature (RFC-Based)¶

Workflow 2: Debugging CI Failure¶

Workflow 3: Creating New Worktree¶

Workflow 4: Refactoring Module¶

Workflow 5: Writing Documentation¶

Pro Tips for @Rules Usage¶

Recommended Workflow (Python Project)¶

Summary: One-Line Mental Model¶

Model Selection Strategy¶

Fast / Mechanical Work¶

Main Coding & Reviews (Daily Driver)¶

Deep Reasoning / Hard Problems¶

High-Risk / Strict Correctness¶

Cursor Auto: How It Works (Practically)¶

How to Influence Auto (What Actually Works)¶

1. Control Scope (Most Important)¶

4. Split Planning and Execution¶

5. Use Composer for Big Tasks¶

Understanding Cursor's Rule Systems¶

Two Separate Rule Systems¶

1. .cursorrules (Root File) - AUTOMATIC ✅¶

2. .cursor/rules/*.mdc Files - MANUAL ON-DEMAND 📋¶

When to Use Each¶

Use .cursorrules (Always Active)¶

Use @Rules *.mdc (On-Demand)¶

Comparison Table¶

Real Examples¶

Example 1: Normal Coding (No @Rules)¶

Example 2: Writing Tests (With @Rules)¶

Example 3: Creating Worktree (With @Rules)¶

Why Have Both Systems?¶

Current Project Structure¶

Prompt Files in .cursor/prompts/¶

Manual Templates (Copy/Paste)¶

Available Prompt Templates¶

debug-ci.txt - CI Failure Debugging¶

design-rfc.md - RFC and Feature Design¶

code-review.txt - Code Review¶

implementation-plan.txt - Implementation Planning¶

How to Use Prompt Templates¶

Subagents and Commands¶

What Are Subagents?¶

Commands vs Subagents¶

Project Custom Subagents (.cursor/agents/)¶

Subagents vs Skills¶

Project-Specific Recommendations¶

1. Always Use @Rules for Specialized Tasks¶

2. Core Rules Are Always Active¶

3. Mandatory Workflow Steps (Enforced by .cursorrules)¶

4. Test Execution Strategy (Use @Rules testing-strategy)¶

5. Documentation Updates (Use @Rules markdown-style)¶

6. Git Worktree Workflow (Use @Rules git-worktree)¶

7. Model Selection for Common Tasks¶

8. Always Reference Project Guidelines¶

Making This Fast (Recommended)¶

Option A: Split View¶

Option B: VS Code / Cursor Snippets (Best UX)¶

Option C: Quick Access to Guidelines¶

Key Takeaways¶

General Principles¶

@Rules System Principles¶

Project-Specific Principles¶

@Rules Quick Reference¶

One-Line Rule¶

Next Steps¶

1. `.cursorrules` (Root File) - AUTOMATIC ✅¶

2. `.cursor/rules/*.mdc` Files - MANUAL ON-DEMAND 📋¶

Use `.cursorrules` (Always Active)¶

Use `@Rules *.mdc` (On-Demand)¶

Prompt Files in `.cursor/prompts/`¶

`debug-ci.txt` - CI Failure Debugging¶

`design-rfc.md` - RFC and Feature Design¶

`code-review.txt` - Code Review¶

`implementation-plan.txt` - Implementation Planning¶

Project Custom Subagents (`.cursor/agents/`)¶