title: Skill Discovery Across AI Coding Platforms authors:

name: Vadim Rudakov email: rudakow.wadim@gmail.com date: ‘2026-05-02’ description: How AI coding platforms discover, index, and load skills — mechanisms, directory layouts, and setup instructions tags:
context_management
architecture options: type: guide birth: ‘2026-04-10’ version: 1.0.1 token_size: 1625

Skill Discovery Across AI Coding Platforms¶

Source and Reproducibility¶

Repository: obra/superpowers Version analyzed: v5.0.7 (commit b7a8f76) Source directory: skills/ — 14 SKILL.md files plus supporting prompt templates Analysis date: April 2026

To reproduce this analysis, clone the repo at this version and examine the skill files directly:

git clone https://github.com/obra/superpowers.git
cd superpowers
git checkout v5.0.7
ls skills/*/SKILL.md

Key files referenced in this analysis:

File	Purpose
`skills/using-superpowers/SKILL.md`	Bootstrap skill — skill invocation rules
`skills/subagent-driven-development/SKILL.md`	Three-role subagent orchestration
`skills/subagent-driven-development/implementer-prompt.md`	Implementer dispatch template
`skills/subagent-driven-development/spec-reviewer-prompt.md`	Spec compliance review template
`skills/subagent-driven-development/code-quality-reviewer-prompt.md`	Code quality review template
`skills/verification-before-completion/SKILL.md`	Verification enforcement
`skills/writing-skills/SKILL.md`	Skill creation methodology
`skills/writing-skills/testing-skills-with-subagents.md`	Adversarial testing methodology
`skills/writing-skills/anthropic-best-practices.md`	Anthropic’s official skill authoring guide
`skills/dispatching-parallel-agents/SKILL.md`	Parallel dispatch pattern
`.opencode/plugins/superpowers.js`	OpenCode plugin — skill path registration
`.claude-plugin/plugin.json`	Claude Code plugin metadata
`.cursor-plugin/plugin.json`	Cursor plugin — skills/agents/commands/hooks paths
`hooks/session-start`	Session bootstrap hook — JSON output injection
`gemini-extension.json`	Gemini CLI plugin metadata
`.codex/INSTALL.md`	Codex installation via symlink
`.github/PULL_REQUEST_TEMPLATE.md`	PR template — contributor requirements
`AGENTS.md` / `CLAUDE.md`	Contributor guidelines for AI agents

How Skills Are Indexed¶

Each platform handles discovery differently. There is no unified indexing layer across platforms:

Platform	Mechanism
Claude Code	Native skill system. Scans `skills/*/SKILL.md`, parses YAML frontmatter, indexes `name` + `description`. Skills are loaded on-demand when the LLM invokes the `Skill` tool.
OpenCode	Plugin pushes `skills/` path into `config.skills.paths`. OpenCode scans that directory, parses frontmatter, and lazy-discovers skills.
Codex	Symlink: `~/.agents/skills/superpowers → skills/`. Codex’s native skill discovery handles the rest.
Cursor	`plugin.json` points to `"skills": "./skills/"`. Cursor’s plugin system discovers them.
Gemini CLI	`gemini-extension.json` with `contextFileName: "GEMINI.md"`. Skills load via `activate_skill` tool.
Qwen Code	Scans `~/.qwen/skills//SKILL.md` and `<project>/.qwen/skills//SKILL.md` — immediate subdirectories only, no nested `skills/` layout. Each skill directory must contain `SKILL.md` with `name` + `description` frontmatter. Symlinks supported (tested: valid dirs, broken links, file targets). Precedence: project > user > extension > bundled.

Superpowers just puts SKILL.md files in a directory with YAML frontmatter and trusts each platform to find them.

The Discovery Chain¶

User input: "help me plan this feature"
         ↓
Platform skills engine searches indexed descriptions
         ↓
Finds: brainstorming SKILL.md description mentions "building features"
         ↓
Loads full SKILL.md content into context (or presents as option)
         ↓
LLM invokes skill via Skill tool

The name becomes the skill identifier (used as superpowers:brainstorming). The description is what the LLM reads to decide whether to invoke the skill.

Token-Efficient Loading¶

From Anthropic’s best practices guide:

“Not every token in your Skill has an immediate cost. At startup, only the metadata (name and description) from all Skills is pre-loaded. Claude reads SKILL.md only when the Skill becomes relevant, and reads additional files only as needed.”

This is a platform-specific guarantee (Claude Code), not universal across all platforms. Once a skill is loaded, it competes with conversation history, code context, and other skills. Skills are aggressively trimmed for token efficiency: <150 words for frequently-loaded, <500 for others.

Setting Up Superpowers Skills for Qwen Code¶

The superpowers repo uses a nested skills/ subdirectory layout (superpowers/skills/<skill-name>/SKILL.md) that Qwen Code cannot discover directly — it expects skills at the immediate level of the base directory.

Recommended layout:

/ai_agents/research/ai_coding_agents/superpowers/          ← git repo (update with git pull)
  skills/
    brainstorming/SKILL.md
    using-superpowers/SKILL.md
    ...

~/.qwen/skills/                                  ← symlinks to each skill
  brainstorming -> /ai_agents/research/ai_coding_agents/superpowers/skills/brainstorming
  using-superpowers -> /ai_agents/research/ai_coding_agents/superpowers/skills/using-superpowers
  ...

Setup commands:

# Clone the superpowers repo into research/
git clone https://github.com/obra/superpowers.git /ai_agents/research/ai_coding_agents/superpowers

# Create symlinks for each skill (run from ~/.qwen/skills/)
cd ~/.qwen/skills/
for skill in brainstorming dispatching-parallel-agents executing-plans \
  finishing-a-development-branch receiving-code-review requesting-code-review \
  subagent-driven-development systematic-debugging test-driven-development \
  using-git-worktrees using-superpowers verification-before-completion \
  writing-plans writing-skills; do
  ln -sfn -r \
    /home/commi/Yandex.Disk/it_working/projects/soviar-systems/ai_engineering_book//ai_agents/research/ai_coding_agents/superpowers/skills/$skill \
    /home/commi/.qwen/skills/$skill
done

Update skills anytime with cd /ai_agents/research/ai_coding_agents/superpowers && git pull — the symlinks follow automatically.