llm-config

Central repository for LLM and agent configuration, reusable skills, prompts, and operating conventions.

Installation

Run the setup script to install agents and skills into Copilot, Codex, or OpenCode user-level directories, or into a project-local configuration directory. Choose one explicit install target; bare ./setup.sh does not perform an install.

./setup.sh --copilot                # symlink Copilot user-level install
./setup.sh --copilot --copy         # copy Copilot user-level install
./setup.sh --copilot-project /path/to/repo
./setup.sh --copilot-project /path/to/repo --copy
./setup.sh --codex                  # symlink Codex skills and agent specs
./setup.sh --codex --copy           # copy Codex skills and agent specs
./setup.sh --opencode               # install OpenCode skills and agents
./setup.sh --opencode --copy        # copy OpenCode skills and agents
./setup.sh --opencode-project /path/to/repo

Copilot install targets:

• Agents → ~/.copilot/agents/*.agent.md
• Skills → ~/.copilot/skills/<name>/SKILL.md

Copilot project install targets:

• Agents → <project>/.github/agents/*.agent.md
• Agents (Claude format) → <project>/.claude/agents/*.agent.md
• Skills → <project>/.github/skills/<name>/SKILL.md
• Instructions → <project>/.github/copilot-instructions.md (generated as a normal file, not a symlink)

Agent frontmatter uses name and target: vscode for Copilot/VS Code discovery. Codex installs the same Markdown agent specs under ~/.codex/agents/; Claude-format agents under .claude/agents/ may need separate Claude-specific metadata, so do not rely on target: vscode as a portable Codex or Claude registration field.

Codex install targets:

• Agent specs → ~/.codex/agents/*.agent.md
• Skills → ~/.codex/skills/<name>/SKILL.md

OpenCode install targets:

• Agents → ~/.config/opencode/agents/*.md
• Skills → ~/.config/opencode/skills/<name>/SKILL.md

OpenCode project install targets:

• Agents → <project>/.opencode/agents/*.md
• Skills → <project>/.opencode/skills/<name>/SKILL.md

OpenCode also discovers skills from Claude-compatible .claude/skills/<name>/SKILL.md and agent-compatible .agents/skills/<name>/SKILL.md paths, but Markdown agent definitions belong in ~/.config/opencode/agents/ globally or .opencode/agents/ per project. The installer writes OpenCode agents as native *.md files with OpenCode frontmatter; the filename becomes the OpenCode agent name.

Restart VS Code or reload the window after running VS Code installs. Restart Codex after running a Codex install. Restart OpenCode after running an OpenCode install.

Per-project usage

To install into a single project with the script:

./setup.sh --copilot-project /path/to/your-project

The installer creates .github/copilot-instructions.md if it does not already exist, and preserves an existing file so you can keep project-specific guidance there.

Codex Usage

To install system-level Codex skills and agent specs:

./setup.sh --codex

The Codex install places this repository's agents under ~/.codex/agents/ and skills under ~/.codex/skills/. By default, these are symlinked back to this repository so edits here are visible after restarting Codex.

With child_agents_md enabled in Codex, the Markdown agent specs in ~/.codex/agents/ can be used as child-agent guidance. Keep reusable methods in skills and focused role definitions in agents.

OpenCode Usage

To install system-level OpenCode skills and agents:

./setup.sh --opencode

The OpenCode install places this repository's skills under ~/.config/opencode/skills/ and agents under ~/.config/opencode/agents/. For a project-local install, use ./setup.sh --opencode-project /path/to/repo.

Structure

• agents/ — agent definitions (*.agent.md with YAML frontmatter)
• skills/ — reusable task skills (SKILL.md with YAML frontmatter, one per directory)
• templates/ — starter templates for new configs
• memory/ — long-lived preferences and conventions
• docs/ — documentation for contributors

Agents

Agent	Description
agent-creator	Custom agent and skill design for this repository
computational-learning-coach	Low-level computational and numerical concepts taught through minimal C++ examples and learning notes
cpp-engineer	Modern C++ design, type safety, ownership, compile-time computation
document-writer	Markdown, Confluence, LaTeX, and report editing
p4-reviewer	Perforce changelist review, submission readiness, changelist comments
performance-engineer	Runtime performance, profiling, bottleneck analysis, C++/CUDA tuning
course-project-worker	Course projects end-to-end: spec extraction, Python ML, RL, LaTeX reports
build-engineer	Boost.Build (b2/bjam), gcc/clang/nvcc error diagnosis, template errors, compilation speed
cmake-project-builder	Modern CMake projects, CMakePresets, C++/CUDA build errors, custom CMake functions
git-integrator	Git branch integration: merge/rebase strategy, conflict resolution, branch sync, and safe recovery
git-publisher	Git commit and push workflow: status, diff review, staging, commit messages, and publishing

Skills (25)

Skill	Domain
agent-customization	Custom agent and skill scope, tools, and delegation wording
coding	General implementation, debugging, refactoring
git-workflow	Git branch workflows, merge/rebase decisions, conflict resolution, history cleanup, and recovery
shell	Bash, zsh, csh/tcsh scripting, terminal automation, pipelines, portability, and shell debugging
boost-build	Boost.Build (b2/bjam), Jamfiles, toolsets, compiler error decoding, compile-time optimization
modern-cmake	Modern CMake, target usage requirements, presets, toolchains, C++/CUDA/nvcc build diagnosis
cpp-elegance	Modern C++ clarity, RAII, concepts, API design
cpp-performance	C++ hot-path optimization, cache, vectorization
computational-learning-notes	Low-level computational, numerical, hardware, and compiler concepts taught through minimal C++ demos
cuda-performance	GPU kernel tuning, shared memory, coalescing
ncu-analysis	Nsight Compute profiling, report extraction, and CUDA bottleneck diagnosis
sparse-linear-algebra	Sparse matrices, iterative solvers, preconditioners, reorderings, graph sparsity
handoff-document	Session handoff files for pausing, resuming, and transferring agent context
p4-review	Perforce code review and changelist comments
p4-workflow	Perforce changelist and shelf management
research	Technical investigation and synthesis
writing	Editing, summarization, style normalization
markdown-editing	Markdown documentation, tables, links, and GitHub formatting
confluence-editing	Confluence pages, team wiki structure, and macro-aware editing
latex-project-build	LaTeX project configuration, engine selection, multi-file compilation, and build diagnostics
latex-tikz	LaTeX documents and TikZ/PGFPlots figures
pdf-requirements-review	PDF spec extraction and project planning
python-ml	Python ML experiment scaffolding
reinforcement-learning	RL experiment design and analysis
report-iteration-overleaf	Report iteration and Overleaf workflows

Conventions

• Keep skills small and composable.
• Prefer one responsibility per file.
• Use Markdown unless a tool requires another format.
• Version changes with clear commit messages.

Adding New Content

1. Copy templates/skill-template.md when creating a new skill.
2. Copy templates/agent-template.md when creating a new agent definition.
3. Document shared assumptions in memory/.
4. Re-run ./setup.sh to pick up new files.