Plugin Skill Create

/plugin-skill-create — Plugin-Convention Skill Scaffold

Generate plugin/skills/<name>/SKILL.md plus eval/case stub plus memory-hook scaffolding. Conforms to:

1. agentskills.io specification — https://agentskills.io/specification (frontmatter, naming, directory layout, progressive disclosure)
2. agentskills.io best practices — https://agentskills.io/skill-creation/best-practices (description style, body length, calibration)
3. agentskills.io description optimization — https://agentskills.io/skill-creation/optimizing-descriptions (trigger phrasing, what+when, keywords)
4. ai-skills plugin conventions — H5 frontmatter trigger pattern, schema-validated frontmatter, <untrusted_content> G1 wrapping where applicable, G7 spawn payload format if the skill spawns subagents

The cached, offline-authoritative form of (1)–(3) is prompt-engineering/skill-authoring-spec.md + prompt-engineering/optimizing-descriptions.md — read those first; re-fetch the live URLs only when network is available.

When to use

• Adding a new workflow skill to the plugin (e.g., a project-specific workflow)
• Adding a knowledge skill to the plugin
• Adding a companion utility skill

Not for: general-purpose Anthropic skill creation — use the upstream skill-creator for that. This scaffold ONLY targets plugin-convention skills under plugin/skills/.

Invocation

/plugin-skill-create new-workflow --invocable --agent-spawn
/plugin-skill-create analyzer --type knowledge
/plugin-skill-create release-notes --type companion --invocable

Arguments

Behavior

1. Validate <name>:
   • Lowercase + hyphens only (per Anthropic skill name convention from docs.claude.com/en/docs/agents-and-tools/agent-skills/best-practices)
   • ≤64 chars
   • Does NOT collide with existing plugin/skills/<name>/
   • If collision: error and suggest --overwrite flag (but this skill never overwrites by default — minimum-change discipline)
2. Generate plugin/skills/<name>/SKILL.md with H5 frontmatter:

---
name: <name>
description: Use this skill when <TODO: trigger phrase> — TODO replace this placeholder describing what the skill does.
context: fork           # if --invocable
argument-hint: "<arg>"  # if --invocable
disable-model-invocation: true  # if --type knowledge AND not --invocable
---

The TODO — token is a hard guard: plugin-skill-audit fails (CRITICAL severity) on any skill description that still contains TODO. This forces the scaffolded description to be edited before merge. Removing the literal TODO from your description is the gate that flips the audit from failing to passing.

3. Generate body skeleton:

   • # /<name> heading
   • Purpose section (placeholder)
   • When to use / Not for sections
   • Invocation examples (placeholder)
   • Arguments table (placeholder)
   • Behavior steps
   • Hard rules
   • Failure modes
   • Memory writes table (if --memory)
   • RALF Loop section (if --ralph)
   • G7 Spawn Payload + Return Contract section (if --agent-spawn)
   • Integration section: rules, hooks, schemas, companions

4. Generate plugin/eval/cases/<name>/eval-case-001.json stub (Tier 3):

{
  "case_id": "<name>-001",
  "skill": "<name>",
  "input": {"prompt": "<test prompt>"},
  "expected": {"contains": ["<expected substring>"]},
  "judge_rubric": "plugin/eval/judge-rubrics/<name>.md",
  "judge_model": "haiku",
  "max_tokens_input": 30000,
  "max_tokens_output": 2000
}

5. Generate plugin/eval/judge-rubrics/<name>.md skeleton with the standard 5-dimension template.

6. Print summary: created files + next steps:

   • Fill placeholder content
   • Run /plugin-skill-audit <name> --strict to confirm spec compliance
   • Run /eval --skill <name> --tier 1

Hard rules

Every scaffolded skill MUST satisfy the agentskills.io specification + ai-skills conventions below. The audit counterpart /plugin-skill-audit enforces the same rules — anything this skill emits MUST pass /plugin-skill-audit <name> --strict.

agentskills.io spec, best-practices, scripts & description rules (cached digest)

The full cached rules live in two digests (read them before scaffolding — single source of truth, progressive disclosure; do NOT re-inline here):

• prompt-engineering/skill-authoring-spec.md — specification (frontmatter, naming, directory layout, 3-level progressive disclosure), best practices, using-scripts, and skill-output eval
• prompt-engineering/optimizing-descriptions.md — the description triggering surface (imperative phrasing, user-intent, be-pushy, trigger eval, 1024 hard limit)

Hard invariants every scaffold MUST satisfy (enforced by /plugin-author audit --strict):

• name 1–64 chars; lowercase a–z, digits, hyphens; no leading/trailing or consecutive (--) hyphens; MUST equal the parent directory name
• description 1–1024 chars; non-empty; what + when; trigger keywords; imperative; no TODO token (audit CRITICAL)
• Body ≤ 5000 tokens AND ≤ 500 lines recommended; file references one level deep from SKILL.md
• Scripts/deps (only when scripts/ is generated): no interactive prompts (agents run non-interactive shells); --help usage; structured data → stdout, diagnostics → stderr; pinned deps or inline PEP 723 / equivalent — see the digest's "Using scripts in skills" section

ai-skills plugin conventions

• Body ≤ 12K chars per project rule (matches the upstream ≤500-line recommendation)
• Description in third person — first-person breaks discovery
• Form A: every description BEGINS with the literal Use this skill when … phrase (ai-skills strict operationalization of upstream imperative phrasing; capability folded into the same sentence)
• Never overwrite existing skill — refuse if plugin/skills/<name>/ already exists
• Plugin-only scope — never scaffolds skills outside plugin/skills/. For project-local skills use upstream skill-creator
• English-only per repo CLAUDE.md
• No absolute user-machine paths in templates or examples

Best practices the scaffold bakes in

Sourced from the digest (prompt-engineering/skill-authoring-spec.md best-practices) — the scaffold output and the author MUST apply, not just cite:

• Start from real expertise — never generate a skill from generic LLM knowledge (produces vague generic procedures). Extract the pattern from a real completed task, or synthesize from project artifacts (runbooks, API specs, code-review comments, VCS history, real failure cases).
• Refine with real execution — the first draft is a draft. Run against real tasks, feed ALL results back (not just failures), read execution traces; wasted steps ⇒ vague/inapplicable/duplicate instructions.
• Design coherent units — one skill = one coherent capability. Too narrow ⇒ many skills per task; too broad ⇒ imprecise activation.
• Output-format templates — emit a concrete template (agents pattern-match well); short inline, long → assets/, referenced.
• Checklists for multi-step workflows; plan-validate-execute for batch/destructive ops (intermediate plan → validate vs source of truth → execute).
• Bundle on reinvention — if traces show the agent re-deriving the same logic each run, ship a tested script in scripts/.

Scripts & dependencies (when the skill ships scripts/)

Sourced from the digest's "Using scripts in skills" section:

• One-off runners over committed code where possible: uvx/pipx run (Python), npx/bunx (Node), deno run (--allow-*), go run. Pin versions (npx eslint@9.0.0).
• Inline deps — Python PEP 723 block + uv run scripts/x.py; pin via PEP 508; requires-python. State runtime-level prerequisites in the compatibility frontmatter field, not body prose.
• Non-interactive (HARD) — no TTY-blocking prompts; input via flags/env/stdin; clear error+usage when a required input is missing.
• --help documented per script; structured output (JSON/CSV/TSV) → stdout, progress/diagnostics → stderr.
• Documented exit codes per failure type; idempotency ("create if not exists"); --dry-run for destructive/stateful ops; predictable output size (harnesses truncate ~10–30K chars — default to a summary, support --offset, or require --output FILE|-).

Failure modes

• Name collision: refuse, suggest different name
• Name fails validation: refuse with examples of valid names (e.g., release-notes, NOT Release Notes or releasenotes)
• Permission denied on write: report path; suggest checking plugin/ writeability

Memory writes

Integration

• Status: internal procedure document for /plugin-author create. Not slash-invocable (frontmatter disable-model-invocation: true). The umbrella reads this file directly when scaffolding a new asset; ai-skills:prompt-engineer pre-reads it before any plugin-asset DEV pass.
• Reachable via: /plugin-author create <name> [--type workflow|knowledge|companion] [--agent-spawn] [--ralph]
• Writes to: plugin/skills/<name>/, plugin/eval/cases/<name>/, plugin/eval/judge-rubrics/<name>.md
• References:
  • prompt-engineering/skill-authoring-spec.md — cached spec + best-practices + scripts + eval digest (offline source of truth)
  • prompt-engineering/optimizing-descriptions.md — cached description-triggering digest
  • agentskills.io specification — https://agentskills.io/specification
  • agentskills.io best practices — https://agentskills.io/skill-creation/best-practices
  • agentskills.io description optimization — https://agentskills.io/skill-creation/optimizing-descriptions
  • Anthropic Agent Skills best-practices — https://docs.claude.com/en/docs/agents-and-tools/agent-skills/best-practices
• Templates: H5 frontmatter pattern, G7 schemas (if --agent-spawn), ralph-budget.md defaults (if --ralph)
• Companion procedure: plugin/skills/plugin-skill-audit/SKILL.md — audit + safe-fix discipline applied by /plugin-author audit.