Agent Skills: OrchestKit Skill Directory

OrchestKit Skill Directory

Dynamic skill discovery: enumerates the installed plugin at runtime so listings are never stale.

CC 2.1.121+ tip: if you just want to find one skill quickly, the native /skills command now has type-to-filter — open it and start typing the skill name. Use /ork:help when you want categorized browsing or rationale for why a skill applies.

Quick Start

/ork:help           # Show all categories
/ork:help build     # Show BUILD skills only
/ork:help git       # Show GIT skills only
/ork:help all       # List every user-invocable skill

Argument Resolution

CATEGORY = "$ARGUMENTS[0]"  # Optional: build, git, plan, quality, memory, config, explore, design, ops, all
# If provided, skip AskUserQuestion and show that category directly.
# $ARGUMENTS is the full string (CC 2.1.59 indexed access)

STEP 0: Dynamic Skill Discovery

ALWAYS run this first to get accurate, up-to-date skill data:

# ${CLAUDE_PLUGIN_ROOT} is set by the plugin runtime and points at the INSTALLED
# plugin. That is the normal case: a marketplace user has no src/ directory.
SKILLS_ROOT = "${CLAUDE_PLUGIN_ROOT}/skills"
matches = Grep(pattern="user-invocable:\\s*true", path=SKILLS_ROOT, output_mode="files_with_matches")

# Dogfooding fallback: inside the OrchestKit repo itself the skills live in the
# source tree. Retry there if the env var was unset or the probe found nothing.
if not matches:
    SKILLS_ROOT = "src/skills"
    matches = Grep(pattern="user-invocable:\\s*true", path=SKILLS_ROOT, output_mode="files_with_matches")

If BOTH probes return zero files, say so plainly ("could not locate the OrchestKit skills directory, checked ${CLAUDE_PLUGIN_ROOT}/skills and src/skills") and stop. Do NOT substitute a remembered list of skill names. Any such list is stale by construction and a confident wrong answer is worse than no answer.

For each matched file, read the frontmatter to get name, description, version, complexity, argument-hint and tags:

Read(file_path=f"{SKILLS_ROOT}/{skill_dir}/SKILL.md", limit=25)

Every number rendered later is derived from this scan, never typed as a literal:

TOTAL = len(matches)   # user-invocable skill count

Build the skill list dynamically. Never hardcode counts or skill names.

STEP 1: Category Selection

If CATEGORY argument provided, skip to STEP 2 with that category.

Otherwise, present categories interactively:

AskUserQuestion(
  questions=[{
    "question": "What type of task are you working on?",
    "header": "Category",
    # 4-option cap (CC schema): every category from STEP 2 is grouped into one of
    # 3 buckets + "Show all". STEP 2 renders the constituent categories for the
    # picked bucket. Descriptions name intents, never skills, so they cannot drift.
    "options": [
      {"label": "Build & ship", "description": "Writing code, tests, git and PRs, UI and design work"},
      {"label": "Plan & assess", "description": "Requirements, planning, quality assessment, review"},
      {"label": "Explore & operate", "description": "Codebase exploration, memory, setup, diagnostics, CI"},
      {"label": "Show all", "description": "List every user-invocable skill"}
    ],
    "multiSelect": false
  }]
)

STEP 2: Render Category

For the selected category, render the skill table from the data gathered in STEP 0.

Category Definitions

Categories are defined by tag predicates, never by a list of skill names, so a newly shipped skill lands in the right bucket without editing this file. Match each discovered skill's frontmatter tags against the sets below (case-insensitive, one hit is enough):

| Category | CLI arg | Matches any of these tags | |----------|---------|---------------------------| | BUILD | build | implementation, feature, testing, coverage, test-generation, verification, e2e | | GIT | git | git, github, commit, pull-request, pr, issue, bug-fix | | PLAN | plan | planning, ideation, prd, requirements, visualization | | QUALITY | quality | quality, assessment, evaluation, code-review, validation, grading | | MEMORY | memory | memory, decisions, graph-memory, consolidation | | CONFIG | config | setup, configuration, onboarding, diagnostics, health-check, dev-loop | | EXPLORE | explore | exploration, codebase, code-search, architecture, discovery | | DESIGN | design | design, design-context, design-tokens, design-to-code, frontend, ui, components, stylecards | | OPS | ops | ci, automation, telemetry, observability, release, migration | | OTHER | (none) | anything the rows above did not match |

A skill matching two categories is listed under both. That is expected, not a bug. OTHER is what makes the totals reconcile: every skill found in STEP 0 must appear somewhere in a full listing, so a skill nobody has categorized yet still shows up.

The STEP 1 picker only offers 3 buckets (the AskUserQuestion schema caps a question at 4 options). Each bucket renders the union of its categories:

| Picker bucket | Renders categories | |---------------|--------------------| | Build & ship | BUILD + GIT + DESIGN | | Plan & assess | PLAN + QUALITY | | Explore & operate | MEMORY + CONFIG + EXPLORE + OPS + OTHER |

For each skill in the category, render:

/ork:{name}  v{version}  {complexity}
  {description}
  Example: /ork:{name} {argument-hint example}

"Show all" — Full Listing

If user picks "Show all", render ALL user-invocable skills grouped by category from STEP 0 data, then close with the derived total: {TOTAL} user-invocable skills. Print TOTAL from len(matches). Never type a number you did not just count.

CC Built-in Commands (2.1.72+)

Not OrchestKit skills — these are Claude Code built-ins:

| Command | Description | Since | |---------|-------------|-------| | /code-review | Review changed code for correctness bugs at a chosen effort level; --comment posts inline PR comments (renamed from /simplify in 2.1.146; the old cleanup-and-fix behavior was removed) | CC 2.1.146 | | /help | Claude Code built-in help | CC 2.1.0+ | | /config | Claude Code configuration | CC 2.1.0+ | | /clear | Clear conversation (preserves background agents) | CC 2.1.72 | | /fast | Toggle fast mode (same model, faster output) | CC 2.1.59+ | | /loop | Recurring interval (e.g. /loop 5m /foo) | CC 2.1.71 | | /plan | Enter plan mode (CC 2.1.222 removed the ultraplan feature, so /ultraplan and the "Refine with Ultraplan" hand-off no longer exist) | CC 2.1.72 | | /team-onboarding | Generate teammate ramp-up guide | CC 2.1.101 | | /recap | Session context restoration after idle | CC 2.1.108 | | /undo | Alias for /rewind | CC 2.1.108 | | /skills | Native skill picker — type to filter | CC 2.1.121 | | /effort | Reasoning effort — sets $CLAUDE_EFFORT env for skills | CC 2.1.72 (env var since 2.1.120) | | /cd | Move session to a new working directory without breaking the prompt cache | CC 2.1.169 | | --safe-mode | Start with ALL customizations disabled (troubleshooting bisect) | CC 2.1.169 (flag/env, not a slash command) |

Performance Tips

• ENABLE_PROMPT_CACHING_1H=1 — Extends prompt cache TTL from 5 min to 1 hour. Set this for long sessions with multi-phase skills (brainstorm, implement, cover). Available on API key, Bedrock, Vertex, Foundry.
• /recap — Restores session context after stepping away. Complements OrchestKit's checkpoint-resume for chain recovery.

Keyboard Shortcuts

| Shortcut | Action | |----------|--------| | Ctrl+F | Find in session output | | Esc | Cancel / dismiss | | Shift+Enter | Newline in input | | Ctrl+C | Cancel operation |

Pro Tip

You don't need to memorize skills. Just describe your task naturally:

"I need to implement user login"     → /ork:implement
"Show me the payment architecture"   → /ork:explore
"Review PR 123"                      → /ork:review-pr
"Is this code good?"                 → /ork:assess
"Plan out the billing redesign"      → /ork:visualize-plan

Related Skills

• /help — Claude Code built-in help
• /ork:doctor — OrchestKit health check
• /ork:setup — Full onboarding wizard