Split 4 core skills to reduce bootstrap token load:
- gh-issues: 865L → 63L SKILL.md + 281L references/phases.md (92%)
- skill-creator: 372L → 65L SKILL.md + 135L references/design-guide.md (82%)
- xurl: 461L → 90L SKILL.md + 192L references/command-reference.md (80%)
- coding-agent: 295L → 125L SKILL.md (57%, no reference file)
Total: 1,993L → 343L SKILL.md + 608L references (~83% bootstrap reduction)
The pattern: keep frontmatter, description, quick-reference tables, and
essential rules in SKILL.md. Move detailed examples, code snippets, and
phase-by-phase instructions to references/ subdirectory. The agent reads
references only when the skill triggers, not at bootstrap.
This follows OpenClaw's existing on-demand loading architecture — descriptions
are in bootstrap, full SKILL.md loads on trigger. This change goes one level
deeper: splitting the SKILL.md itself so trigger-time loads are also lighter
for the 90% case where the quick reference is sufficient.
384de62b0c