Add deployment guidance and missing frontmatter fields to skill-creator

[flound1129]

Problem

The skill-creator guides users through creating and iterating on skills but never tells them where to place the finished skill so Claude Code actually discovers it. This is the most common reason skills fail to trigger after creation — they end up in the working directory instead of a discovery path.

The skill-creator also only teaches 3 frontmatter fields (name, description, compatibility) while Claude Code supports 11. Users who need disable-model-invocation, context: fork, or allowed-tools have no guidance within the skill-creation workflow.

What this PR changes

Adds a "Deploying the finished skill" section to the skill-creator SKILL.md covering:

• The four skill placement levels (enterprise, personal, project, plugin) with priority ordering
• Project root clarification (.claude/skills/ lives at project root)
• Monorepo nested discovery behavior
• How the name field relates to directory name (optional, defaults to dir)
• A deployment checklist
• All supported frontmatter fields with behavioral notes (context:fork turns skill body into subagent prompt, effort:max is Opus-only, hooks links to docs)
• String substitutions ($ARGUMENTS, ${CLAUDE_SKILL_DIR}, etc.)
• Dynamic context injection (!command syntax)
• Updated core loop summary to include the deployment step

All content sourced from the official Claude Code docs at code.claude.com/docs/en/skills and platform.claude.com/docs/en/agents-and-tools/agent-skills/.

Methodology

We downloaded and cross-referenced the official Anthropic skill docs against the skill-creator SKILL.md and identified every gap where the skill-creator is silent on information users need to successfully deploy a skill.

Two rounds of code review caught issues in the initial draft:

• A fabricated "name must match directory" constraint (softened to "recommended, not enforced" per the official docs)
• Missing enterprise tier in the placement table
• Missing effort:max Opus-only qualifier
• Missing context:fork behavioral explanation
• Missing hooks docs pointer
• Missing monorepo nested discovery

Adversarial pressure testing

We ran 3 pressure scenarios, each tested once without any skill guidance (baseline) and once with the updated guidance that documents the same placement/frontmatter information this PR adds:

Metric	Baseline (no skill)	With skill guidance
Correct discovery path	1/3	3/3
Correct directory structure	1/3	3/3
Correct frontmatter fields (no hallucinations)	0/3	3/3
Correct priority ordering	0/1	1/1
Hallucinated fields/variables	4	0
Used deprecated commands system	1	0

Test 1 (manual-invoke skill): Baseline hallucinated trigger: manual. With guidance, agent used correct disable-model-invocation: true.

Test 2 (skill in ~/Downloads not triggering): Baseline diagnosed using the deprecated .claude/commands/ system and flat files. With guidance, agent correctly identified ~/.claude/skills/api-linter/SKILL.md.

Test 3 (monorepo + subagent + dynamic injection): Baseline hallucinated agent: true, tools:, and $GIT_BRANCH fields, and got priority ordering backwards. With guidance, agent used correct context: fork + agent: Explore and correct priority (enterprise > personal > project).

Environment

Harness	Version	Model	Model ID
Claude Code	2.1.71	Claude Opus 4.6	claude-opus-4-6