coding-skills

Canonical Anthropic Agent Skills for software engineering — packaged in a portable, harness-agnostic format.

Each skill is a self-contained folder with SKILL.md (frontmatter + body), optional scripts/, and optional reference/. Skills are consumable by Claude Code, SkillNet, Cursor 2.0 (via MDC frontmatter), OpenAI Codex Skills, and the Microsoft Agent Framework with no transformation.

Claude Code consumers also get parallel-friendly agents under agents/ (e.g. @code-reviewer, @security-auditor, @architect-review) and slash commands under commands/ (e.g. /coding-skills:review, /coding-skills:tdd, /coding-skills:commit) that wrap the skills — same source of truth, richer harness ergonomics. Other harnesses ignore the agents/ and commands/ directories.

Skill catalogue

See INDEX.md for the full list. Two kinds:

• Workflow skills — what to do (verbs): running-tdd-cycles, reviewing-changes, designing-architecture, creating-block-issues (Spec Kit projects — one issue per tasks.md PR-stack block), committing-changes (also installs an optional PR-size CI gate that fails PRs over 1000 changed lines, excluding tests/docs/lockfiles/generated), implementing-blocks (Spec Kit projects — one PR-stack block end-to-end).
• Rule skills — conventions to apply (nouns): python-conventions, go-conventions, solidity-conventions, shell-discipline, engineering-philosophy.

Workflow skills cross-reference rule skills; agents activate the rule skill alongside the workflow skill when the file or language matches.

Frontmatter convention

Every SKILL.md uses a hybrid of Anthropic Skills (required) + Cursor MDC (optional, for auto-activation in Cursor):

---
# Anthropic Skills
name: python-conventions
description: Apply Python project conventions ... Use when writing or reviewing Python code ...
allowed-tools: Read, Bash(uv *), Bash(uv run ruff *), ...

# Cursor MDC (ignored by Anthropic, used by Cursor)
globs: "**/*.py"
paths: "**/*.py"
---

Fields explicitly excluded from canonical SKILL.md frontmatter:

• model: — Claude-Code-only; original agent model class is preserved as a one-line header in reference/<agent>.md.

Install

Claude Code (plugin)

Inside any Claude Code session:

/plugin marketplace add swell-agents/coding-skills
/plugin install coding-skills@swell-agents

Then /reload-plugins. All 11 skills become available (auto-activated by description), plus 3 parallel-review agents and 7 slash commands:

Claude Code (manual)

Clone the repo and symlink individual skills into ~/.claude/skills/.

SkillNet

skillnet download https://github.com/swell-agents/coding-skills/tree/main/skills/<skill-name>

Cursor

Add the repo path to your project's MDC source roots; the globs: / paths: frontmatter drives auto-activation.

Bootstrap a project

Skills are description-matched at retrieval; that is good enough for most cases but not guaranteed. When a project should always apply the engineering principles — not just when the matcher fires — patch the project's instruction file once:

In Claude Code:

/coding-skills:bootstrap

In any other harness:

bash ~/.claude/plugins/swell-agents/coding-skills/scripts/bootstrap.sh
# or, after `git clone`:
bash scripts/bootstrap.sh

The script is patch-only. It never creates instruction files; it only appends the engineering-skills block to ones that already exist.

• `CLAUDE.md` — if present, the script appends the engineering-skills reference block (idempotent via the <!-- coding-skills-bootstrap --> marker). If missing, no-op.
• `AGENTS.md` — never modified. Present is fine, absent is fine. The plugin does not opine on which instruction file is canonical; AGENTS.md as canonical, CLAUDE.md as canonical, or a mix is all acceptable.
• `.cursorrules` — if the file exists, the same engineering-skills block is appended. The script does not create it.
• `.cursor/rules/` — if the directory exists, no-op (you maintain rule files directly).

If none of those files exists, the script prints a hint and exits 0 — create whichever instruction file fits your project (per the rubric's R2: presence of any one is the bar), then re-run.

Re-running once the markers are in place is a no-op.

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.