jleechanorg/jleechan-skills

description: Claude Skills & Commands - Modular Agent Skills Library for Claude Code, Antigravity, Codex, and Cursor
type: llm-orchestration
execution_mode: immediate
---
🧠 Agent Skills
Portable SKILL.md packages for debugging, evidence, research, reviews, and
agent workflows. Skills are the source of truth; slash commands are optional
shortcuts.
---
Executive Summary
Six highlighted skills form a portable quality loop around implementation.
Shift left to challenge decisions and prove the problem before code is trusted;
shift right to locate failures, audit evidence, and prevent repeat escapes.
Highlighted skills · left/right shift strategy
| Shift left · before the code | Shift right · after the code |
|---|---|
| **/superpowers-quick** — jleechanorg's autonomous wrapper around Superpowers by Jesse Vincent; design before writing code. | **/4layer** — unit → end-to-end → MCP/HTTP → browser; the first reproducing layer owns the investigation. |
| **/advice + /web-advice** — fan out independent reviewers before committing to an approach. | **/evidence-review** — require every claim to resolve to a real, provenance-backed artifact. |
| **/redgreen** — require a test you watched fail before accepting the bug or its fix as real. | **/harness** — when a failure repeats, fix the process that allowed it through. |
The six highlighted workflow entries are the jleechanorg /superpowers-quick
wrapper, /advice + /web-advice, /redgreen, /4layer, /evidence-review,
and /harness. These prompts and packages are portable and do not depend on
WorldAI or private infrastructure.
> **Third-party attribution:** /superpowers-quick
> is a jleechanorg-authored autonomous wrapper around
> Superpowers, which is made by Jesse
> Vincent. Superpowers itself is **not
> authored, owned, or maintained by Jeffrey Lee or jleechanorg**.
Strategy source: “Shift left, shift right — a portable skill system”.
---
Intelligent self-setup (any platform)
Ask your coding agent to inspect the repository and install only what it needs:
For a deterministic install, use the bundled installer. It installs complete
skill packages (including their helper files), commands, agents, and scripts
under CLAUDE_HOME (default: ~/.claude). It refuses a nonempty target by
default; --backup preserves that target before installing and --merge
updates it in place. A routine merge preserves active files whose names collide
with archived packages; after reviewing those warnings, opt into their
migration with --merge --migrate-archives.
This is an isolated smoke test: inspect "$INSTALL_ROOT/claude/skills" and
remove only that exact temporary directory when finished. For an install into
your real ~/.claude, first read the backup and rollback notes in
INSTALL.md.
See INSTALL.md for target locations, isolated verification, and
safe rollback.
🏛️ How It Works: Skills First, Thin Slash Pointers
1. **Skills (.claude/skills/) are the source of truth** — protocols, schemas, checklists, and helper scripts live in standard SKILL.md folders.
2. **Slash commands (.claude/commands/) are thin pointers** — no separate logic, just an ergonomic shortcut that loads and runs the skill.
3. **Cross-agent portability** — because every skill follows the SKILL.md spec, it runs natively in Claude Code, Antigravity, Codex, and Cursor without translation.
Read a skill before using it: some skills require a local binary, browser
session, or integration that is not bundled with this export.
An opt-in curated portable edition provides native-tool
derivatives with a hash manifest and isolated installation. Canonical Claude
packages retain their existing behavior; file installation, runtime discovery,
and successful workflow execution are verified separately.
---
📋 Skills at a Glance
The catalog below links to the canonical skills. Commands are convenience
pointers and are intentionally not the documentation target.
| Command | Full name / skill | When to use it / expected outcome |
|---|---|---|
| /advice (/smart-advisor) | advice | Use at a design or implementation decision. Give the question plus a PR, ref, or path; independent reviewers read the real scope and return a synthesized recommendation. For a PR gate, incomplete or truncated review context can only yield WITHHELD, never approval. |
| /ablation (/abal) | ablation | Use when diagnosing a bug in LLM behavior or stochastic systems where the cause is disputed. Manipulates the real captured input and replays it through the real production path, one variable per arm, with a verbatim control arm run first. |