Reusable agent skills I use to make coding agents more useful, less vague, and less likely to sprint confidently into a wall.

These skills are small, composable workflows for OpenCode and other SKILL.md-compatible agents. They cover planning, frontend design, Cloudflare work, codebase exploration, repository walkthroughs, Jira automation, tmux-managed servers, and skill authoring.

The repo is intentionally boring: each skill lives in its own folder at the repository root, with a SKILL.md file and any supporting scripts or references next to it.

Point OpenCode at this repo by adding it to your opencode.json:

{
  "skills": {
    "paths": ["~/path/to/skills"]
  }
}

OpenCode scans custom skill paths for **/SKILL.md, so top-level folders like tmux/SKILL.md and spec-planner/SKILL.md are picked up automatically.

If you need to refresh a runtime copy manually:

cp -R ~/path/to/skills/* ~/.config/opencode/skills/

Do not author new skills there. Author them in this repo, then sync or configure OpenCode to load this repo directly.

Agents are powerful, but they fail in familiar ways:

• They start building before understanding the problem.
• They produce generic plans that sound good and survive zero contact with the codebase.
• They run long-lived processes inline and turn the terminal into soup.
• They miss project context, naming conventions, and architectural pressure points.
• They need a repeatable way to use specialized references without bloating every prompt.

These skills make those failure modes explicit and give the agent better rails.

• grill-me — Interview the user relentlessly about a plan or design until the decision tree is resolved.
• spec-planner — Develop implementation-ready specs through skeptical questioning, discovery, drafting, and refinement.
• repo-walkthrough — Explain how a repository works at a calibrated level, from novice overview to detailed architecture tour.

• tmux — Manage dev servers, watchers, logs, and other long-running commands in tmux panes.
• using-git-worktrees — Create isolated git worktrees for feature work without trampling the current workspace.
• diff-walkthrough — Produce guided walkthroughs of diffs, PRs, MRs, commit ranges, or staged changes with configurable depth and output format.
• improve-codebase-architecture — Find refactoring opportunities that deepen modules and improve testability and agent navigability.
• index-knowledge — Generate hierarchical AGENTS.md knowledge files for a codebase.

• cloudflare — Work across Cloudflare Workers, Pages, KV, D1, R2, Durable Objects, Workers AI, Vectorize, networking, security, and IaC.
• jira-tool — Create, update, transition, and manage Jira tickets through the bundled CLI workflow.
• frontend-design — Apply React-oriented UI/UX guidance for component structure, layout, accessibility, and interaction design.

• build-skill — Create, review, and validate effective skills with the expected structure and progressive disclosure patterns.
• install-skill — Install agent skills from GitHub repositories using gh.
• librarian — Research library internals, code patterns, and architecture across GitHub, npm, PyPI, and crates repositories.

Each skill should look like this:

skill-name/
  SKILL.md
  references/
  scripts/

Only SKILL.md is required. Put large references and executable helpers next to the skill instead of bloating the main instructions.

Every SKILL.md should include frontmatter:

---
name: skill-name
description: Use this when the agent should do a specific thing.
---

Skill names should be lowercase kebab-case.

MIT. Use these, change them, delete the spicy bits, make them yours.