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.
See INDEX.md for the full list. Two kinds:
- Workflow skills — what to do (verbs):
running-tdd-cycles,reviewing-changes,designing-architecture,managing-github-issues,committing-changes(also installs an optional PR-size CI gate that fails PRs over 1000 changed lines, excluding tests/docs/lockfiles/generated). - 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.
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 inreference/<agent>.md.
Inside any Claude Code session:
/plugin marketplace add swell-agents/coding-skills
/plugin install coding-skills@swell-agents
Then /reload-plugins. All 10 skills become available (auto-activated by description), plus 3 parallel-review agents and 6 slash commands:
| Slash command | Wraps |
|---|---|
/coding-skills:review [scope] |
Three @code-reviewer + @security-auditor + @architect-review agents in parallel; aggregates into one Quality Gate Summary |
/coding-skills:commit [scope] |
committing-changes skill |
/coding-skills:tdd [requirement|phase] |
running-tdd-cycles skill |
/coding-skills:pm <plan|start|next|advance|status|create-issues> |
managing-github-issues skill |
/coding-skills:design [topic] |
designing-architecture skill |
/coding-skills:bootstrap |
One-shot wiring into the project's instruction file |
Clone the repo and symlink individual skills into ~/.claude/skills/.
skillnet download https://github.com/swell-agents/coding-skills/tree/main/skills/<skill-name>
Add the repo path to your project's MDC source roots; the globs: / paths: frontmatter drives auto-activation.
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.shThe script detects every instruction file the project uses (CLAUDE.md, AGENTS.md, .cursorrules) and appends a coding-skills reference block to each. If none exist, it creates CLAUDE.md. Idempotent — re-running once the marker is in place is a no-op.
See CONTRIBUTING.md.
MIT — see LICENSE.