55 reusable Claude Code skills with eval coverage across 46 skills. Each skill is a focused workflow you can invoke with /skill-name in any Claude Code session.
What Are Skills?
Skills are Claude Code plugins that provide specialized capabilities. They're stored in skills/golem-powers/ and can be installed into any project. Each skill has a trigger (slash command), a description, and focused instructions for a specific workflow.
Skill Categories
Note: the live, always-current skills inventory is the auto-generated /golems/skills page; this table is a curated overview.
Skills are AI-agnostic — each skill can include adapters for different AI CLI tools:
CLI
Adapter
Claude Code
adapters/claude.md
Cursor
adapters/cursor.md
Gemini CLI
adapters/gemini.md
Codex CLI
adapters/codex.md
Kiro
adapters/kiro.md
Installing Skills
Skills are available as a Claude Code plugin:
# From your project directoryclaude --plugin-dir ~/Gits/golems/skills/golem-powers
Or symlink into your project's .claude/commands/ directory for per-project access.
Eval Coverage
46 of 55 skills have eval suites with structured assertions. Evals verify that skills trigger correctly and produce the expected behavior — and that they beat baseline: recent A/B-evaled skills include /cursor-multitask (baseline 78.6% → with-skill 100%), /fleet-wrap (60% → 100%), and /phoenix-human-view (45.8% → 100%), with raw scoring committed in each skill's evals/results/. See individual skill pages for eval details.
Recent Skill Updates (2026-05-17)
Friction-reduction sprint Phase 5 — seven skill PRs merged in one night, each backed by measured signals (not hypothesis).
/large-plan Phase N+1 — every plan producing code or config now requires a separate evaluator subagent (≥8/10 or ITERATE). Self-audit explicitly does not satisfy /goal. Closes the "self-grade as evaluator" loophole observed in 4 of 4 /goal outputs the night of 2026-05-17. PR #407.
/orc C4 parallel-by-default + /agent-routing auto-dispatch triggers — batch reads ≥3, transcription ≥2, or web research ≥1 now fan out parallel sub-agents in the same turn, no permission ask. Live eval flipped Agent calls from 0 → 4 across both orc and coach launcher matrices. PR #405.
/cmux-agents envelope-vs-delivery pairing — any [FROM=X TO=Y TYPE=Z] envelope in the author's pane must be paired with send_to_agent in the same turn. Field evidence: 64 orphan envelopes in one Codex session. Also adds auto-workspace-categorization for new_split panes. PR #409.
/session-handoff inherited-citation suspect rule — after consuming a handoff doc or resuming post-compaction, all inherited file:line citations are SUSPECT until re-Read. Fabricated cites get logged with compaction-fabrication tag. Defense-in-depth against post-compaction pattern-completion fabrication. PR #410.
/presentation-builder standing-preference inline (canonical in /coach) — repeated stylistic asks (≥2 in a session) like "less terms, more visuals" become STANDING for the rest of the session. Self-prompt + self-check before every deliverable. Slide-deck eval: 9 correction turns → 2, slide 2 word count 377 → 17. PR #404.
Hyphen-aware launchers in ralph #7 — skill-creatorClaude, maakaf-homeClaude, 6pm-miniClaude and friends now resolve, so mcp__cmux__spawn_agent({repo:"skill-creator"}) works.
