Skills

Small, composable agent skills for real engineering work. Each skill is independent, trigger-activated, and works with any Agent Skills spec-compatible harness — Claude Code, pi, opencode, and anything skills.sh-compatible.

Why

Monolithic agent frameworks (GSD, BMAD, Spec-Kit) try to own your entire workflow. They work until they don't — then debugging the framework is harder than doing the work yourself.

These skills take the opposite approach:

• Small — each skill does one thing well (diagram, debug, grill, handoff)
• Composable — skills reference each other when useful, never depend on each other
• Adaptable — fork, modify, keep what works, discard what doesn't
• Model-agnostic — plain markdown instructions, no vendor lock-in
• Harness-agnostic — Agent Skills spec frontmatter, same SKILL.md for every runtime

Install

Any Agent Skills harness (recommended)

npx skills@latest add TheAkshitS/skills

Pick the skills and agents you want. To grab one skill non-interactively:

npx skills@latest add TheAkshitS/skills -s c4-views -y

Claude Code

bash scripts/link-skills.sh

Symlinks every skill into ~/.claude/skills. Re-runnable; safe to overwrite.

pi

bash scripts/install-pi.sh link            # symlink into ~/.pi/agent/skills

Also symlinks into ~/.claude/skills and ~/.agents/skills in the same pass. Two more install modes (no-symlink alternatives) and full troubleshooting are in docs/pi-install.md, the canonical recipe.

Manual / drop-in

The repo ships committed mirror copies under .claude/skills/ and .agents/skills/ (see Mirror in the glossary). Copy either folder into your harness's skill directory and the skills just work.

Skills

• c4-views — Generate and review C4 model architecture diagrams in Mermaid, Structurizr DSL, or C4-PlantUML. Trigger: "diagram", "architecture", "C4", "system context", "container diagram", "component diagram"
• external-model — Run a prompt through a different AI model via the opencode, cursor-agent, or kiro-cli command-line tools. Trigger: "second opinion", "what would GPT-5/Gemini say", "delegate to cursor/opencode/kiro", "run with model X", "ask another model"

Harness support

Harness	Install	Slash command	Notes
Claude Code	bash scripts/link-skills.sh	/c4-views, /external-model	~/.claude/skills
pi	bash scripts/install-pi.sh link	/skill:c4-views, /skill:external-model	~/.pi/agent/skills
opencode	drop into ~/.agents/skills	/skill:c4-views	same as pi global fallback
skills.sh (installer)	npx skills add TheAkshitS/skills	depends on host	installer, not a client — picks the host harness

All four rows read the same SKILL.md; only the install path and slash-command prefix differ. Many more harnesses support the format — Claude, Cursor, Codex, Gemini CLI, GitHub Copilot, Goose, Roo Code, Amp, and others. See the official Agent Skills client showcase for the full list. This repo documents install for the four it tests against.

Skill structure

skills/<skill-name>/
├── SKILL.md           # Instructions loaded by the agent (required)
├── AGENTS.md          # Skill-specific agent hints (optional)
├── README.md          # Standalone distribution docs (optional)
├── references/        # Long-form material loaded on demand (optional)
└── scripts/           # Utility scripts (optional)

Skills stay flat under skills/ until 3+ skills exist, then organize into buckets: engineering/, productivity/, misc/. See docs/repo-layout.md for details.

Develop locally

Symlink every skill into all supported harnesses' global skill dirs:

bash scripts/link-skills.sh

Validate frontmatter and mirror sync before committing:

bash scripts/validate-skills.sh

Refresh the in-repo distribution mirrors after editing a skill:

bash scripts/sync-copied-skills.sh

Releases

Automated via release-please on push to main. See docs/local-workflow.md for the file layout, semver bump rules, and bootstrap notes.

License

MIT — see LICENSE.