Tasks:

Collect parameters — Scope: create | refine, Type: simple | standard | complex, Depth: base | extended | full
Read frontmatter.md — Discovery metadata, trigger patterns
Read structure.md — Folder layout gated by Type
Read depth.md — LOC limits, nesting gated by Depth
(complex) Read scripting.md — Automation standards
Capture requirements — purpose, triggers, outputs
Invoke skill-summarizer with skill style-standards — Extract voice, formatting, taxonomy
Invoke deep-research — Domain research for skill topic
Plan with 3 agents — file inventory, section structure, content framework
Execute per Scope:
- (create) Author new artifacts; select template:
  - simple - DEFAULT
  - standard
  - complex
- (refine) Compare input to existing frontmatter; see refine.md:
  - Input = existing → optimize (density, fixes, quality)
  - Input > existing → upgrade (expand structure or depth)
  - Input < existing → downsize (combine, refactor, remove low-relevance)
Validate — Quality gate, LOC compliance, structure match

Dependencies:

deep-research — Domain research via parallel agents
skill-summarizer — Voice and formatting extraction (with skill style-standards)
report.md — Sub-agent output format

[REFERENCE]: index.md — Complete file listing

[1][FRONTMATTER]

Dictum: Metadata enables discovery before loading.

Frontmatter indexed at session start (~100 tokens). Description is ONLY field parsed for relevance—quality determines invocation accuracy.

Guidance:

Discovery — LLM reasoning matches description to user intent. No embeddings, no keyword matching.
Trigger Density — Include file types, operations, "Use when" clauses. Every word aids matching.
Voice — Third person, active, present tense. Prohibit: 'could', 'might', 'probably', 'should'.

Best-Practices:

Length — 1-2 sentences. Concise triggers outperform verbose explanations.
Classification — Include type and depth fields for refine workflow detection.

[2][STRUCTURE]

Dictum: Type determines breadth—folder existence defines capability scope.

Type gates folder creation. Structure defines WHAT exists; Depth constrains HOW MUCH content.

[INDEX]	[TYPE]	[FOLDERS]
[1]	Simple	SKILL.md only
[2]	Standard	+index.md, references/, templates/
[3]	Complex	+scripts/

Guidance:

Naming — Skill folder matches frontmatter name exactly. Kebab-case throughout.
Index — Standard/Complex require index.md at root listing all reference files.
Upgrade Path — Start with simplest type satisfying requirements.

Best-Practices:

Directory Purpose — references/ for domain knowledge, templates/ for output scaffolds, scripts/ for automation.
File Limit — Max 7 files in references/ (including nested).

Dictum: Depth determines comprehensiveness—hard caps prevent bloat.

Depth enforces LOC limits and nesting rights. Each level adds +50 SKILL.md, +25 reference files (cumulative).

[INDEX]	[DEPTH]	[SKILL.MD]	[REF_FILE]	[NESTING]
[1]	Base	<300	<150	Flat only
[2]	Extended	<350	<175	1 subfolder
[3]	Full	<400	<200	1-3 subfolders

Guidance:

Nesting Gate — Subfolder requires 3+ related files OR distinct domain concern.
Content Scaling — Base: 1-2 items per Guidance/Best-Practices. Extended: 2-4. Full: comprehensive.
LOC Optimization — Density over deletion; see depth.md§LOC_OPTIMIZATION.
Content Separation — SKILL.md = WHY, references = HOW; see depth.md§CONTENT_SEPARATION.

Best-Practices:

Dictum: Deterministic automation extends LLM capabilities.

Complex type enables scripts/ folder for external tool orchestration, artifact generation, validation.

Guidance:

Justification — Script overhead demands explicit need: tool wrapping, exact reproducibility, schema enforcement.
Depth Scaling — Base/Extended: single script. Full: multiple when distinct concerns justify.

Best-Practices:

Type Selection — Standard suffices for most skills. Complex only when automation is core purpose.
Augmentation — Scripts support workflows; core logic remains in SKILL.md and references.

Dictum: Templates enforce canonical structure.

Templates define output scaffolds. Agent combines user input with template skeleton for consistent artifacts.

Guidance:

Best-Practices:

Dictum: Gates prevent incomplete artifacts.

[VERIFY] Completion:

[REFERENCE] Operational checklist: →validation.md