Backend 421 installs

skill-creator

by starchild-ai-agent/official-skills

Scaffold new skills with valid frontmatter, directory layout, and a starter SKILL.md. Use when building a new reusable workflow or wrapping a new API (e.g.…

Skill content

Core Principles

Concise is key. The context window is a shared resource between the system prompt, skills, conversation history, and your reasoning. Every line in a SKILL.md competes with everything else. Only add what you don't already know - don't document tool parameters visible in the system prompt, don't prescribe step-by-step workflows for things you can figure out. Focus on domain knowledge, interpretation guides, decision frameworks, and gotchas.

Progressive disclosure. Skills load in three levels:

- Always in context - name, emoji, and description appear in <available_skills> in every conversation. This is how you decide which skill to activate. The description must be a strong trigger.

- On activation - the full SKILL.md body is loaded via read_file when you decide the skill is relevant. This is where workflow, guidelines, and decision trees live.

- On demand - scripts/, references/, and assets/ are only loaded when explicitly needed. Heavy content goes here, not in the body.

This means: keep the SKILL.md body lean (< 500 lines). Put detailed API docs in references/. Put automation in scripts/. The body should be what you need to start working, not an encyclopedia.

Degrees of freedom. Match instruction specificity to task fragility:

- High freedom (text guidance) - When multiple approaches are valid. Write natural language explaining WHAT and WHY, not step-by-step HOW. Example: "Check funding rates and social sentiment to gauge market mood."

- Medium freedom (pseudocode + params) - When a preferred pattern exists but details can vary. Describe the approach with key parameters. Example: "Use RSI with period 14, buy below 30, sell above 70."

- Low freedom (scripts in scripts/) - When operations are fragile, require exact syntax, or are repetitive boilerplate. Put the code in standalone scripts that get executed, not loaded into context. Example: Chart rendering with exact color codes and API calls.

Default assumption: you are already smart. Only add context you don't already have.