~/.agents — central agent-skills hub

All skills for AI agents (Cursor, Codex, Claude Code, Hermes, etc.) live here, in one place. This folder is both the on-disk working directory and a public mirror: github.com/dlgrv/agents.

Add skills to a project

ln -sfn ~/.agents <project>/.agents

Symlink the whole ~/.agents directory — the project gets all its contents (skills, README, etc.) and stays always up to date.

Why symlink the whole directory

• Single source of truth: update something in the hub — it's instantly current in every project.
• Cursor reliably reads .agents/skills/ only inside a project; global user skills are unstable for it.
• Zero sync commands, zero cron jobs, zero diverging copies.

Structure — real copies, attribution in README

skills/<name>/ are real committed copies (not symlinks, not submodules) of only the used subfolder of each upstream repo. We don't vendor whole upstream monorepos (they can be tens of thousands of files) — just the skill we use, pinned to the upstream commit it was taken from (see the attribution table). This survives upstream deletion by design, needs no git submodule update --init on a fresh machine, and keeps the repo at a few MB.

~/.agents
├── skills/                      ← real dirs, committed files
│   ├── karpathy-guidelines/           ← from multica-ai/andrej-karpathy-skills
│   ├── improve-codebase-architecture/ ← from mattpocock/skills
│   ├── thermo-nuclear-code-quality-review/ ← from cursor/plugins
│   ├── shadcn/                        ← from shadcn-ui/ui
│   ├── plan-eng-review/               ← from garrytan/gstack
│   ├── sepia/                         ← from Nanako0129/sepia
│   ├── hallmark/                      ← from Nutlope/hallmark
│   ├── anti-slop/                     ← from rand/cc-polymath
│   ├── no-ai-slop/                    ← from petergyang/no-ai-slop
│   ├── i-have-adhd/                   ← from ayghri/i-have-adhd
│   ├── anthropic-frontend-design/     ← from anthropics/skills
│   ├── design-taste-frontend/, high-end-visual-design/,
│   │   redesign-existing-projects/, image-to-code/  ← from Leonxlnx/taste-skill
│   ├── omo-frontend/                  ← local build from code-yeongyu/oh-my-openagent (see below)
│   ├── drawio-skill/                  ← from Agents365-ai/drawio-skill
│   └── (14 × superpowers skills)      ← from obra/superpowers
├── subagents/                   ← own subagents, unified Cursor + Claude Code (see below)
├── cursor/rules/                ← Cursor-specific global rule (see below)
├── scripts/
│   └── update-skill.sh          ← refresh one skill from upstream
├── README.md                    ← this file
└── LICENSE                      ← MIT (for my files only; skills keep upstream licenses)

Attribution

Third-party skills belong to their authors; this repo ships only the skill subfolders, each pinned to the upstream commit it was copied from. _ pins = the upstream commit SHA the copy was taken from (also recorded in skills/<name>/.upstream-pin).

Skill	Upstream (we use only the skill subfolder, not the whole repo)	License (upstream)	_ pin
karpathy-guidelines	multica-ai/andrej-karpathy-skills skills/karpathy-guidelines	MIT	2c60614
improve-codebase-architecture	mattpocock/skills skills/engineering/improve-codebase-architecture	see upstream	6654f6b
thermo-nuclear-code-quality-review	cursor/plugins cursor-team-kit/skills/thermo-nuclear-code-quality-review	see upstream	bdf7aa3
shadcn	shadcn-ui/ui skills/shadcn	see upstream	683a5a9
plan-eng-review	garrytan/gstack plan-eng-review	see upstream	ad84005
sepia	Nanako0129/sepia skills/sepia	MIT	4c8d782
hallmark	Nutlope/hallmark skills/hallmark	MIT	13ac0ec
anti-slop	rand/cc-polymath skills/anti-slop	MIT	baa2df1
no-ai-slop	petergyang/no-ai-slop skills/no-ai-slop	MIT	000650b
i-have-adhd	ayghri/i-have-adhd skills/i-have-adhd	MIT	4092de0
superpowers (14 skills)	obra/superpowers skills/<name>	MIT	b36e082
anthropic-frontend-design	anthropics/skills skills/frontend-design	see upstream	41bbe19
design-taste-frontend	Leonxlnx/taste-skill skills/taste-skill	see upstream	ccbc156
high-end-visual-design	Leonxlnx/taste-skill skills/soft-skill	see upstream	ccbc156
redesign-existing-projects	Leonxlnx/taste-skill skills/redesign-skill	see upstream	ccbc156
image-to-code	Leonxlnx/taste-skill skills/image-to-code-skill	see upstream	ccbc156
omo-frontend (local build)	code-yeongyu/oh-my-openagent — SKILL.md из packages/skills-loader-core/.../frontend, references из packages/shared-skills/skills/frontend/references (design/designpowers/perfection); Layer A taste-файлы — симлинки на скиллы из этого хаба; брендовые Layer B (Open Design) — заглушки	see upstream	local
drawio-skill	Agents365-ai/drawio-skill skills/drawio-skill	MIT	cfe6131
improve-agent-harness-token-efficiency	X: @ericzakariasson — промпт из твита (Cursor learnings), не репо; зеркало telegra.ph	see author	tweet (transcribed)

Update a skill from upstream

cd ~/.agents
./scripts/update-skill.sh <skill-name> <subpath-in-upstream-repo>
# e.g.:
./scripts/update-skill.sh karpathy-guidelines skills/karpathy-guidelines
git add skills/<name> && git commit -m "update <name> from upstream"

The script shallow-clones the upstream repo and rsyncs the subfolder over the inlined copy. Check the diff before committing — upstream may have moved ahead.

Add a new third-party skill

1. Clone the upstream repo (shallow), copy the skill subfolder into skills/<name>/.
2. Write the pinned upstream SHA into skills/<name>/.upstream-pin.
3. Add a row to the attribution table above and commit.

Your own materials

Personal files (subagents, rules, prompts, notes) go in as regular repo files — no submodules. The catalog grows horizontally without rework.

Subagents (subagents/) — unified across agents

Own subagents in the common Claude-Code format (md + YAML frontmatter). Cursor and Claude Code both read this format; each is wired to the single source with symlinks:

~/.agents/subagents/          ← symlinked WHOLE into ~/.cursor/agents and ~/.claude/agents
├── bulk-reader.md          ← reads files, returns only a summary; composer-2.5, readonly
└── boilerplate-writer.md   ← tests/config/stubs by pattern; composer-2.5

# Cursor (user scope, all projects — whole-dir symlink)
ln -sfn ~/.agents/subagents ~/.cursor/agents
# Claude Code (user scope, all projects — whole-dir symlink)
ln -sfn ~/.agents/subagents ~/.claude/agents

Pattern: token-saving delegation (inspired by Spotify's shunt / Portal AiKA modes) — subagents keep bulky file contents out of the main thread. Note: the model: field like composer-2.5 is Cursor-specific — Claude Code ignores unknown ids and uses its default.

Cursor rule (cursor/rules/)

cursor/rules/delegation.mdc — Cursor-specific .mdc rule telling the main agent to route 3+ file reads / 350+ line files to the subagents above. There is no cross-agent standard for global rules, so this stays Cursor-only:

# whole-directory symlinks (one command per app; new files appear automatically)
ln -sfn ~/.agents/subagents       ~/.cursor/agents   # same dir doubles as ~/.claude/agents
ln -sfn ~/.agents/subagents       ~/.claude/agents
ln -sfn ~/.agents/cursor/rules    ~/.cursor/rules
# per project (documented rules location, guaranteed effect)
mkdir -p <project>/.cursor/rules && ln -sfn ~/.agents/cursor/rules/delegation.mdc <project>/.cursor/rules/delegation.mdc

~/.cursor/rules/ is not an officially documented rules source — for a guaranteed effect, paste the rule body into Settings → Rules → User Rules (one-time) or symlink it into a project's .cursor/rules/.