IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE. YOU MUST USE IT.

This is not negotiable. This is not optional. You cannot rationalize your way out of this. </EXTREMELY-IMPORTANT>

Auto-Update Check (First Run)

# Quick update check (run silently)
cd ~/.claude/skills 2>/dev/null || cd "$(readlink -f ~/.claude/skills)" 2>/dev/null
if git fetch origin --dry-run 2>&1 | grep -q .; then
    git pull origin main --quiet
fi

If updates were pulled, briefly note: "Radisha skills updated to latest version."

This ensures you always have the latest skills. Skip this check on subsequent skill invocations in the same session. </IMPORTANT>

Language Rule

• User writes in Serbian → Respond in English
• User writes in German → Respond in English
• User writes in any language → Respond in English

This ensures consistency and readability across all documentation, plans, code comments, and conversations. </IMPORTANT>

Core Principles

Simplicity First - Make every change as simple as possible. Impact minimal code.

No Laziness - Find root causes. No temporary fixes. Senior developer standards.

Minimal Impact - Changes should only touch what's necessary. Avoid introducing bugs.

Every Plan Leads to a Pull Request - Keep PR reviewability in mind from the start.

Emoji Policy - Avoid emojis in code, comments, and general output. Exception: User-facing interactive menus may use emojis for visual clarity (e.g., ✅ Continue, ⬅️ Revert).

Quick Commands

Use these shortcuts for fast access to common skills:

For /plan specifically: Present an interactive menu using AskUserQuestion to ask which planning type (feature/bugfix/refactor/docs/architecture), then invoke the selected skill. </IMPORTANT>

How to Access Skills

In Claude Code:

1. Use the Skill tool to invoke skills by name (e.g., planning-feature)
2. When you invoke a skill, its content is loaded and presented to you—follow it directly
3. Fallback: If skills are not found in the project, manually read from ~/.claude/skills/ using the Read tool

In Cursor:

1. Use @skill-name mention in Chat/Composer
2. Skills are loaded from .cursor/skills/ or user skills folder

In other environments: Check your platform's documentation for how skills are loaded.

Platform Tool Mapping

Different platforms use different tool names. Use this mapping:

Using Skills

The Planning Rule

Planning comes BEFORE implementation. Always.

Use Cursor's Plan Mode for planning phases. After completing the plan:

1. Switch to Agent Mode
2. Immediately save the plan to planning/ folder
3. Then begin implementation

This ensures plans persist in files and can be resumed later. </IMPORTANT>

Choose the right planning skill:

• planning-feature - New features (includes changelog + test case prompts)
• planning-bugfix - Bug fixes (asks about changelog)
• planning-refactor - Improving existing code (best practices, testability, reduce duplication)
• planning-docs - Documentation (no changelog)
• review-architecture - Architecture documentation (iterative, chapter-by-chapter, no code)

Debugging skills:

• debugging-rocprof-sys - Interactive debugging of rocprofiler-systems issues (logs, gdb, strace)

Exploration skills (before planning):

• exploration-explore-code - Systematic exploration of unfamiliar code before extraction

Programming skills by language:

• C++: programming-cpp, programming-cpp-design-patterns, programming-cpp-stl-algorithms, programming-cpp-naming-rules
• Python: programming-python
• CMake: programming-cmake-best-practices

Code quality skills:

• code-smells - Detect anti-patterns and refactoring opportunities (used by pr-review)
• pr-review - Review PRs or local changes; produces structured report
• pr-review-interactive - Walk through PR review findings one by one, accumulate inline comments in a PENDING review, submit at end

Library-specific skills:

• AMD SMI: libraries-amd-smi - GPU/CPU monitoring and management

Radisha management skills:

• radisha-update - Update radisha to latest version
• radisha-help - Full workflow reference (this skill)

Async coordination skills:

• watch - Schedule a recurring poll of an external condition (PR merge, CI green, Jira state, background build); runs a follow-up action when met and self-stops

After implementation, create test plan:

• Invoke testing-testplan skill
• Create planning/testplan-<name>.md with verification scenarios

After test plan, offer unit tests:

• For C++: Invoke testing-gtest-gmock skill
• For Python: Invoke testing-pytest skill
• Write tests ONE BY ONE, waiting for user approval after each test

After tests, offer Pull Request:

• Invoke git-prepare-pull-request skill
• Create PR with Motivation, Technical Details, Test Plan

Workflow:

1. Receive user request
2. Determine task type (feature / bugfix / refactor / documentation)
3. If task is non-trivial → invoke appropriate planning-* skill
4. Assess PR scope - split into multiple PRs if > 800 lines
5. Create task list based on planning (use platform's task tracking tool)
6. Invoke domain-specific skills (programming, documentation, testing)
7. Execute with task tracking
8. After each step: ask for validation (see below)
9. Mark each completed task in both task tracker AND the plan file
10. Create test plan - planning/testplan-<name>.md
11. Ask about unit tests after implementation
12. Ask about creating PR after tests

Step-by-Step Validation (REQUIRED)

Use the AskUserQuestion tool for interactive clickable menus instead of text-based options. </IMPORTANT>

After each step:

1. Autonomous verification (if code was changed):
   • Run relevant tests if they exist
   • Check for linter errors
   • Verify the change works as expected
   • Note any issues found
2. Show a summary of what was implemented:
   • Brief description of changes
   • Key files changed
   • Verification results (tests passed/failed, etc.)
   • Any decisions made or assumptions
3. Use AskUserQuestion tool with these options:

{
  "title": "Step Validation: [Step Description]",
  "questions": [{
    "id": "validation",
    "prompt": "I've completed [step]. Please review the changes above.",
    "options": [
      {"id": "continue", "label": "✅ Continue - Proceed to next step"},
      {"id": "revert", "label": "⬅️ Revert & Stop - Revert changes and stop"},
      {"id": "improve", "label": "🔄 Improve - Try a different approach"}
    ]
  }]
}

Validation flow:

Complete Step N
      │
      ▼
Autonomous verification (run tests, check logs)
      │
      ▼
Show summary + verification results
      │
      ▼
AskUserQuestion (interactive menu)
      │
      ├── "continue" ──────→ Mark step done, proceed to Step N+1
      │
      ├── "revert" ────────→ Revert changes, stop implementation
      │
      └── "improve" ───────→ Revise implementation, ask again

Example:

First, show the changes with verification:

Completed: Step 2 - Add validation to user input

Changes:

• Added validate_input() function in src/utils/validation.cpp
• Added input checks in process_request()
• Returns std::optional<error> on validation failure

Verification:

• ✅ All existing tests pass
• ✅ No linter errors
• ✅ Manual check: invalid input correctly rejected

Then use AskUserQuestion tool for the interactive menu.

Re-planning When Things Go Wrong

When to stop and re-plan:

• Multiple unexpected errors or failures
• The approach reveals unforeseen complexity
• Tests fail in ways that suggest the design is wrong
• You find yourself making "just one more fix" repeatedly

How to re-plan:

1. Stop current implementation
2. Document what went wrong and what was learned
3. Switch back to Plan Mode
4. Create a revised plan incorporating the new understanding
5. Get user confirmation before resuming

Subagent Strategy

Use subagents liberally to keep the main context window clean and focused.

When to use subagents:

• Research and exploration tasks
• Parallel analysis of multiple files/components
• Complex problems that benefit from more compute
• Tasks that can be isolated and delegated

Subagent rules:

• One task per subagent - keep execution focused
• Offload research - don't clutter main context with exploration
• Parallel analysis - spin up multiple subagents for independent investigations
• Clear handoff - provide subagent with all necessary context upfront

Example use cases:

• "Explore how authentication works in this codebase" → subagent
• "Find all usages of deprecated API" → subagent
• "Analyze performance of these 3 modules" → 3 parallel subagents

Elegance Check (Non-Trivial Changes)

When to apply:

• Changes that affect multiple files
• New abstractions or patterns being introduced
• Refactoring existing code
• Architectural decisions

Skip this for:

• Simple, obvious fixes
• One-line changes
• Typo corrections
• Config updates

If a fix feels hacky:

"Knowing everything I know now, is there a more elegant solution?"

Challenge your own work before presenting it to the user.

When Unclear: Offer Options

If you're not sure which skill applies to the user's request, use AskUserQuestion tool:

{
  "title": "Task Type",
  "questions": [{
    "id": "task_type",
    "prompt": "Which best describes what you need?",
    "options": [
      {"id": "feature", "label": "New feature - Add new functionality"},
      {"id": "bugfix", "label": "Bug fix - Fix something that's broken"},
      {"id": "refactor", "label": "Refactoring - Improve existing code"},
      {"id": "docs", "label": "Documentation - Create or update docs"},
      {"id": "question", "label": "Just a question - No action needed"}
    ]
  }]
}

Do NOT guess. When in doubt, ask.

When Uncertain About Solution: Ask User

Ask when:

• Multiple valid approaches exist with different trade-offs
• You're unsure if the solution fits the project's architecture
• The requirement is ambiguous
• The solution has significant implications (performance, breaking changes, etc.)
• You're making assumptions that could be wrong

How to ask:

"I have a few options for this. Before I proceed:

Option A: [Description] - [Trade-offs] Option B: [Description] - [Trade-offs]

Which approach would you prefer? Or should I consider something else?"

Or for confirmation:

"I'm planning to [approach]. This assumes [assumption]. Is this correct, or would you prefer a different approach?"

Do NOT:

• Assume you know best when multiple valid solutions exist
• Make significant architectural decisions without confirmation
• Proceed with uncertainty when a quick question could clarify

Questions Without Actions

If the user is asking a question that doesn't require any code changes, file modifications, or implementation:

• Use ask skill for explanations, clarifications, or informational requests
• No planning is needed
• No TodoWrite is needed
• Simply answer the question directly

Examples of "ask" scenarios:

• "What does this function do?"
• "How does authentication work in this project?"
• "Explain the difference between X and Y"
• "What's the best practice for...?"

The Skill Rule

Invoke relevant or requested skills BEFORE any response or action. Even a 1% chance a skill might apply means that you should invoke the skill to check. If an invoked skill turns out to be wrong for the situation, you don't need to use it.

digraph skill_flow {
    "User message received" [shape=doublecircle];
    "Is it a question only?" [shape=diamond];
    "ask skill" [shape=box];
    "Is type clear?" [shape=diamond];
    "Offer skill options" [shape=box];
    "Is task non-trivial?" [shape=diamond];
    "Determine task type" [shape=diamond];
    "planning-feature" [shape=box];
    "planning-bugfix" [shape=box];
    "planning-refactor" [shape=box];
    "planning-docs" [shape=box];
    "Create TodoWrite from plan" [shape=box];
    "Might domain skill apply?" [shape=diamond];
    "Invoke domain Skill tool" [shape=box];
    "Execute and mark done" [shape=box];
    "Ask about unit tests" [shape=box];
    "Respond" [shape=doublecircle];

    "User message received" -> "Is it a question only?";
    "Is it a question only?" -> "ask skill" [label="yes"];
    "ask skill" -> "Respond";
    "Is it a question only?" -> "Is type clear?" [label="no, needs action"];
    "Is type clear?" -> "Offer skill options" [label="unclear"];
    "Offer skill options" -> "Is task non-trivial?";
    "Is type clear?" -> "Is task non-trivial?" [label="clear"];
    "Is task non-trivial?" -> "Determine task type" [label="yes, >2 steps"];
    "Is task non-trivial?" -> "Might domain skill apply?" [label="no, trivial"];
    "Determine task type" -> "planning-feature" [label="new feature"];
    "Determine task type" -> "planning-bugfix" [label="bug fix"];
    "Determine task type" -> "planning-refactor" [label="refactoring"];
    "Determine task type" -> "planning-docs" [label="documentation"];
    "Determine task type" -> "review-architecture" [label="architecture doc"];
    "review-architecture" -> "Respond";
    "planning-feature" -> "Create TodoWrite from plan";
    "planning-bugfix" -> "Create TodoWrite from plan";
    "planning-refactor" -> "Create TodoWrite from plan";
    "planning-docs" -> "Create TodoWrite from plan";
    "Create TodoWrite from plan" -> "Might domain skill apply?";
    "Might domain skill apply?" -> "Invoke domain Skill tool" [label="yes"];
    "Might domain skill apply?" -> "Execute and mark done" [label="no"];
    "Invoke domain Skill tool" -> "Execute and mark done";
    "Execute and mark done" -> "Ask about unit tests";
    "Ask about unit tests" -> "Respond";
}