Editing Rules and Skills

Two flavors of work — classify first

Before touching anything, identify which kind of work this is. The two flavors look superficially similar but want very different voices and processes, and classifying wrong is the single biggest way this skill gets misapplied.

• Correction-driven — Josh has been burned by Claude getting something wrong and wants the lesson encoded so it doesn't recur. Reactive: a frustration arrived from another project or session. Load correction-driven.md before editing.
• Informative / capability — Josh wants to surface a tool, command, workflow, or process so Claude can use it competently. Proactive: no incident, just a capability worth documenting. Load informative-capability.md before editing.

If unsure, ask Josh in one sentence: "Is this you correcting a behavior, or you wanting me to know about a capability?"

Most existing global skills (github, bin-scripts, brainstorming, planning, dispatch-parallel-agents, humanizer) are informative. Most of the flat behavioral rules are correction-driven.

When this skill fires

• Editing the flat behavioral rules in claude/CLAUDE.md
• Editing the code-authoring or code-review standards in the agent defs (claude/agents/implementer.md, claude/agents/code-reviewer.md)
• Editing project rules, which live as sections in the project CLAUDE.md
• Editing or creating any skill file (global or project)
• Adding new items to any of the above
• Anytime Josh says "let's document this so it doesn't happen again" (correction) or "let's make sure you know about X" (informative)

Where it lands

Voice is one axis; location is a separate one. Whichever voice the work takes, it also needs a home: which file a behavioral lesson lands in (the flat list or an agent def) and which tree a rule or skill belongs in (global everywhere, or project only in this repo). Get either wrong and the artifact pollutes every project, silently stops applying, or lands in the wrong voice. Load placement.md for both decisions.

What this skill is NOT

• Not a license to make the terse behavioral rules or agent defs dramatic for the sake of drama.
• Not a license to invent incidents to anchor a correction.
• Not a license to change the semantics of a rule while reframing it.
• Not a license to dress informative reference up in correction voice.