madarauchiha-314/sherma/tasks

• Create a skill that coding agents like claude, cursor etc can use to understand sherma and all the constructs that it provides
• this skill will be used by consumers of the sherma framework/library to create agents using the constructs that sherma provides
• the skill should be compatible with claude skills and https://agentskills.io/home
• when given an abmiguous request to build and agent, sherma skill should ask follow up questions so that it has all the context needed to create the agent
• balance between asking follow up questions and making decisions

Chat Iterations

Iteration 1: Initial implementation

Implemented the plan as written — created skills/sherma/ with SKILL.md, skill-card.json, references (8 doc copies), and assets (4 example copies). Also created .claude/skills/sherma/SKILL.md as a separate file.

Iteration 2: Avoid duplicate SKILL.md

User feedback: "do we need to write in .claude folder, aren't we already creating a skills/sherma folder at the root?"

Decided to keep both locations since .claude/skills/sherma/SKILL.md is required for Claude Code's /sherma slash command to work, while skills/sherma/ is the agentskills.io format. But the content should not be duplicated.

Iteration 3: Symlink the whole folder

User feedback: "the whole folder should be symlinked, not just the SKILL.md"

Changed approach: instead of symlinking just SKILL.md, symlinked the entire directory:

• .claude/skills/sherma → ../../skills/sherma (directory symlink)

This means the single source of truth is skills/sherma/ and Claude Code reads it through the symlink. The SKILL.md file has both agentskills.io frontmatter (name, description, license) and Claude Code frontmatter (user-invocable, allowed-tools, argument-hint) merged into one file.

Git stores symlinks as the path they point to, so this works when checked into GitHub.