Skip to main content

Extensible Skills

The customization layer. SpecWeave builds on Claude Code's native skill system with a clean separation:

  • Core instructions live in SKILL.md, stable and versioned
  • Your project-specific rules live in skill-memories/*.md

Skills self-load their customizations through plain LLM instructions embedded in SKILL.md — a cross-platform approach that tells the skill to read your preferences before executing. No shell execution is required. A typical instruction looks like:

Skill Memories: If .specweave/skill-memories/{skill-name}.md exists, read and apply its learnings.

This replaces the previous shell-based dynamic context injection (DCI) approach for skill memories. DCI (Claude Code's !`command` syntax) is still used for other purposes like loading skill context or PR diffs, but skill memories now use instruction-based loading for better cross-platform compatibility.

The result: you correct Claude once ("use React Hook Form, not useState for forms"), and that preference is applied automatically in every future session.

In This Section​

Extensible Skills Specification​

The formal specification — extensibility category definitions, dynamic context injection specification, detection algorithm, and conformance requirements.

Implementation Guide​

Practical how-to — getting started, architecture, cascading lookup, real-world examples, skill memory format, and FAQ.

Customization Without Modification​

Overview page with quick reference table and links to the specification and guide.

Claude Skills Deep Dive​

How skills work under the hood — progressive disclosure architecture, comparison with other AI tool systems, and the evolution from prompts to programs.

Self-Improving Skills (Reflect)​

The Reflect system auto-learns from corrections and saves them to skill memories. Correct once, applied forever.

Development Guidelines​

Design principles for skill authoring — how to design skills that users can extend without modification.

See Also​