skill-creator

name	skill-creator
description	Guide for creating effective skills that extend different agents' capabilities. Use when creating new skills or updating existing skills with specialized knowledge, workflows, or tool integrations.
version	2.4.0
kind	hybrid
triggers	["create a new skill","add a skill"]
intent	meta
config_dir	~/.config/skill-config/skill-creator
guardrails	["Classify the skill before writing it. Guidance skills get zero scripts; pipeline skills keep judgment out of code.","A script must never dump raw data into agent context — emit a projection plus a handle to the full artifact.","SKILL.md is a router. Detail belongs in references/, loaded on demand.","A skill update that changes behavior updates its prompts/ too, if it has one — stale examples are worse than none."]
resources	["<SKILL_PATH>/scripts/scaffold.sh","<SKILL_PATH>/references/frontmatter.md","<SKILL_PATH>/references/script-contract.md","<SKILL_PATH>/references/guidance-skills.md","<SKILL_PATH>/references/data-app-skills.md","<SKILL_PATH>/references/webapp-list-view.md","<SKILL_PATH>/references/webapp-card-board.md","<SKILL_PATH>/prompts/create-data-app-skill.md","<SKILL_PATH>/prompts/migrate-to-data-app-skill.md"]
tools	["bash"]
created_at	2026-05-30T00:00:00.000Z
updated_at	2026-07-30T00:00:00.000Z

Skill Creator

Read references/house-conventions.md first. It carries the path, tooling, deployment, and working-style conventions of this setup — a skill that ignores them is valid and still feels foreign. Optional config: ~/.config/skill-config/skill-creator/skill.properties (default_author, auto_stow_after_creation), protocol in references/frontmatter.md.

Step 0 — Classify before writing

Every skill is one of three kinds, declared as kind: in frontmatter. The kind determines the entire shape of the skill. Choosing wrong is the most expensive mistake available here, because it is discovered only after the skill is written.

kind	test	shape
guidance	Encodes judgment. Nothing to compute at runtime.	Prose only. Zero scripts. Router SKILL.md + references/.
pipeline	Operates on real data — files, repos, APIs, metrics.	Thin SKILL.md + scripts/ + a declared output contract.
hybrid	Both, in cleanly separable sections.	Declare which sections are which. Suspect it is really two skills.

If you cannot name the concrete runtime input a script would read, it is guidance. Adding scripts to a guidance skill produces rigid output that is wrong in every situation the author did not foresee.

The boundary rule

Do not ask "can this be scripted." Ask whether the operation is total or a judgment.

• Total — exactly one right answer for every input: counting, parsing, filtering, sorting, diffing, aggregating, schema validation, formatting. Always a script. An agent doing this is slow, expensive, and wrong at the edges.
• Judgment — depends on context: relevance, ranking, causation, what to recommend, what to tell the user, and what to do when a script fails. Always the agent. Freezing judgment into code yields output that is confidently wrong wherever the author's guess ran out.

Corollary: tunable values live in data, the logic applying them lives in code. Thresholds, weights, and limits belong in a JSON or properties file beside the script, never inline. Retuning must be an edit to data, not to logic.