| name | using-lrt |
| description | Use when starting any conversation - establishes how to route tasks to the right skills using the routing table |
If you were dispatched as a subagent to execute a specific task, skip this skill.
Task Routing
Match the user's task to a row and follow the route. Read the context files, then invoke the skills in order.
| Work area | Context (read first) | Skills | Skip |
|---|
| Building TheRock | directory-map.md, docs/workflows/building.md (plugin) | lrt-rocm:the-rock | tasks/, design docs |
| Building via rocm-systems | directory-map.md, docs/workflows/building.md (plugin) | lrt-rocm:hip-ocl-monorepo-build | tasks/, design docs |
| Building on Windows | directory-map.md, docs/workflows/building.md (plugin) | lrt-rocm:pal-rocr-windows-build | tasks/, design docs |
| Debugging test failures | docs/workflows/debugging.md (plugin) | lrt-rocm:systematic-debugging, then lrt-rocm:regression-bisect-hip-ocl if regression | unrelated source trees |
| Fixing a bug | docs/workflows/debugging.md (plugin) | lrt-rocm:systematic-debugging -> lrt-rocm:test-driven-development | unrelated source trees |
| Implementing a feature | task file in tasks/active/, docs/workflows/feature-development.md (plugin) | lrt-rocm:brainstorming -> lrt-rocm:writing-plans -> lrt-rocm:subagent-driven-development | unrelated source trees |
| Reviewing code | docs/workflows/review-and-pr.md (plugin) | lrt-rocm:stage-review -> lrt-rocm:process-review | build output |
| Preparing a PR | docs/workflows/review-and-pr.md (plugin) | lrt-rocm:prep-pr, lrt-rocm:squash-prep | build output |
| Build system changes | docs/adding-third-party-dep.md (plugin), CMakeLists.txt | lrt-rocm:the-rock (for context) | test output |
| Submodule coordination | directory-map.md, .gitmodules | rk.py for topic/branch management | build output |
If the user explicitly requests a skill by name (e.g., /lrt-rocm:the-rock), invoke it directly — no routing needed.
If no row matches, respond normally without invoking skills.
digraph routing_flow {
"User message received" [shape=doublecircle];
"Explicit skill requested?" [shape=diamond];
"Invoke requested skill directly" [shape=box];
"Match task to routing table row" [shape=box];
"Row matched?" [shape=diamond];
"Read files from 'Context' column" [shape=box];
"Invoke skills from 'Skills' column (in order)" [shape=box];
"Announce: 'Using [skill] to [purpose]'" [shape=box];
"Follow skill exactly" [shape=box];
"Respond normally (no skill needed)" [shape=doublecircle];
"User message received" -> "Explicit skill requested?";
"Explicit skill requested?" -> "Invoke requested skill directly" [label="yes"];
"Invoke requested skill directly" -> "Follow skill exactly";
"Explicit skill requested?" -> "Match task to routing table row" [label="no"];
"Match task to routing table row" -> "Row matched?";
"Row matched?" -> "Read files from 'Context' column" [label="yes"];
"Read files from 'Context' column" -> "Invoke skills from 'Skills' column (in order)";
"Invoke skills from 'Skills' column (in order)" -> "Announce: 'Using [skill] to [purpose]'";
"Announce: 'Using [skill] to [purpose]'" -> "Follow skill exactly";
"Row matched?" -> "Respond normally (no skill needed)" [label="no"];
}
Instruction Priority
- User's explicit instructions (CLAUDE.md, direct requests) — highest priority
- LRT skills — override default system behavior where they conflict
- Default system prompt — lowest priority
How to Access Skills
Use the Skill tool. When you invoke a skill, its content is loaded and presented to you — follow it directly. Never use the Read tool on skill files.
Skill Types
Rigid (TDD, debugging, verification): Follow exactly. Don't adapt away discipline.
Flexible (patterns, workflows): Adapt principles to context.
The skill itself tells you which.
