Feature: Allow skill authors to mark intermediate text output as collapsible

Summary

Skill/command authors currently have no way to make intermediate assistant text output collapsible in the terminal UI. Tool call results automatically get the condensed treatment (showing last few lines with "ctrl+o to expand"), but plain text output between tool calls is always shown in full.

Use case

When building multi-step skills, the intermediate "thinking out loud" text between tool calls creates significant visual noise.

These messages are useful for debugging but clutter the final output. The user primarily cares about the final summary.

Desired behavior

Provide a mechanism for skill authors to mark sections of assistant text output as collapsible/condensable, similar to how tool call results are automatically treated. This could be:

1. A special markdown annotation or tag in the skill prompt that instructs the UI to collapse intermediate text
2. A frontmatter option in the skill definition (e.g., collapse_intermediate: true)
3. A way to wrap text output in a collapsible block that defaults to showing the last few lines with "ctrl+o to expand"

Current workarounds

The only current option is to delegate all work to a subagent, which hides intermediate steps entirely. This loses the real-time step-by-step visibility that users appreciate — a collapsible middle ground would be ideal.

Environment

• Claude Code CLI (also relevant to desktop app, web app, and IDE extensions)
• Affects any skill/command that performs multi-step orchestration