Will's custom Agent Skills — portable, model-agnostic reference guides that coding agents (Claude Code, Codex, Copilot CLI, Gemini CLI, …) load on demand to apply proven techniques.
Each skill is a directory containing a SKILL.md (metadata + instructions) plus optional references/, scripts/, and assets/. Agents read the description to decide when to load a skill, then pull in references and scripts only as needed. See the Agent Skills format for details.
| Skill | What it covers | Use when |
|---|---|---|
| windows-shell-execution | Running commands on/through Windows — cmd.exe, Windows PowerShell 5.1 vs PowerShell 7+, Win32-OpenSSH, and WSL. Multi-layer quoting/parsing model, the PowerShell native-arg-passing version matrix (incl. the 7.3+ revert-to-Legacy trap), sshd cmd /c mechanics, host keys, CRLF, and CLIXML noise. Every load-bearing claim verified live on Windows 11. |
A command that works on Linux mangles quotes/drops args on Windows, Host key verification failed, /bin/bash^M: bad interpreter, PowerShell stripping quotes from JSON, or reaching WSL over SSH. |
| claude-model-routing | Choosing which Claude model (Opus/Sonnet/Haiku) and effort level to run a task — or assign a subagent — for the best quality per token. The model × effort decision (raise effort before jumping a tier), routing by difficulty adjusted for stakes/reversibility, cascade/escalate-on-failure, the orchestrator + worker subagent pattern, and the cost truth that total cost = price × turns-to-correct-completion. Numbers verified 2026-06-27 against official pricing, benchmarks, and the routing literature. | Picking a model for a task, dispatching a subagent/Task/Workflow, deciding whether to escalate to Opus or drop to Haiku, or "which model should I use for this?" |
| codex-model-routing | Choosing a Codex-selectable GPT-5.6 Sol/Terra/Luna or GPT-5.4 Mini configuration by expected cost to correct, verified completion. Uses a task matrix and dated configuration-level cost, quality, runtime, and token evidence; treats Luna as the cost center, Terra as a runtime/token or max bridge, Sol as the judgment ceiling, and enforces a hard four-agent delegation cap including guarded ultra use. |
Picking a Codex model or effort, deciding whether to escalate or down-route, assigning a bounded Mini worker, controlling model-routing cost, or preventing runaway recursive subagents. |
- Claude Code / Codex / Copilot CLI / Gemini CLI: clone or symlink a skill directory into your runtime's skills location (e.g.
~/.claude/skills/or~/.agents/skills/), then invoke it by name. - Directly: point any agent at a skill's
SKILL.mdand ask it to follow it.
<skill-name>/
├── SKILL.md # required: YAML frontmatter (name, description, …) + instructions
├── references/ # optional: detailed docs loaded on demand
├── scripts/ # optional: runnable helpers/diagnostics
└── assets/ # optional: templates, data, images
MIT © Will Sarg