Writing
Notes from the runtime.
Long-form pieces about agent runtimes versus raw APIs, what makes a "skill" a useful unit of knowledge, and where skillterm sits between shell-gpt-style one-shot tools and full agent stacks.
-
Tab completion is the cheap half: how skillterm decouples expensive generation from interactive lookup
Generating a skill costs 30–120 seconds and real API money. Pressing Tab has to be free. skillterm splits the two by writing SKILL.md to disk at generation time and reading it at Tab time. Here is how the split is structured and why it matters.
shell-integrationperformancearchitecture -
Anatomy of a SKILL.md: what skillterm actually writes when it generates a skill
A walkthrough of the SKILL.md format skillterm uses — YAML frontmatter, markdown body, optional references and scripts — and why the Claude Agent Skills convention is the right artefact for a CLI assistance layer.
skill-formatclaude-skillsarchitecture -
Agent runtime vs. raw LLM API: why skillterm calls Claude Code instead of the Anthropic API
The skillterm README is loud about one decision: invoke a headless agent runtime rather than the LLM API. Here is the engineering argument, the costs that decision avoids, and the trade-offs it accepts.
architectureclaude-codeagent-runtime