yskills

Agent Skills I actually use, open-sourced one at a time.

Both skills solve the same problem from different angles: you run many agent sessions in parallel, you come back hours later, and every one of them greets you with a blank prompt.

They follow the Agent Skills open format, so they work in Claude Code, Claude.ai, Codex, Cursor, OpenCode, Goose, and any other skills-compatible agent.

Which one do I want? human-context-rebuild tells you where you are; whats-next tells you where you are and hands you the next move as clickable options (plus a verdict on whether the session can simply be closed).

Install

Claude Code — plugin (recommended)

Installs both skills, keeps them updatable, and survives a machine rebuild.

/plugin marketplace add lycfyi/yskills
/plugin install yskills@yskills

Or without entering the REPL:

claude plugin marketplace add lycfyi/yskills
claude plugin install yskills@yskills

Plugin skills are namespaced by the plugin, so invoke them as /yskills:whats-next and /yskills:human-context-rebuild. Claude also triggers them automatically when you say things like "我们到哪一步了" or "what should I do next".

Update later with /plugin marketplace update yskills then /plugin update yskills.

Claude Code — plain files

If you'd rather not use plugins, drop the skill folders in and get the shorter /whats-next command:

curl -fsSL https://raw.githubusercontent.com/lycfyi/yskills/main/install.sh | bash

That clones to a temp dir and copies both skills into ~/.claude/skills/. From a clone:

git clone https://github.com/lycfyi/yskills.git && cd yskills
./install.sh                      # both skills -> ~/.claude/skills/
./install.sh whats-next           # just one
./install.sh --link               # symlink instead of copy, so edits track the repo

Or by hand — a skill is just a folder with a SKILL.md:

mkdir -p ~/.claude/skills
cp -r skills/whats-next skills/human-context-rebuild ~/.claude/skills/

Codex

Codex reads skills from ~/.agents/skills (user level) and .agents/skills (repo level):

curl -fsSL https://raw.githubusercontent.com/lycfyi/yskills/main/install.sh | bash -s -- --agent codex

From a clone, ./install.sh --agent codex, or manually:

mkdir -p ~/.agents/skills
cp -r skills/* ~/.agents/skills/

Restart Codex so it picks up the new skill metadata.

Claude.ai, Claude Desktop, Cowork, and cloud sessions

These don't read ~/.claude/skills on your machine — you upload the skill to your account instead. Build the zips:

scripts/package.sh
# dist/whats-next.zip
# dist/human-context-rebuild.zip

Then upload each .zip under Settings → Capabilities → Skills on claude.ai (or Customize in the Claude Desktop sidebar). Skills enabled on your account are what Cowork and scheduled cloud sessions load at session start.

Both skills stick to the six spec-allowed frontmatter fields, so they upload without the Unexpected key(s) in SKILL.md frontmatter error that Claude Code-only fields cause.

Cursor, OpenCode, Goose, and other agents

./install.sh --agent cursor        # ~/.cursor/skills
./install.sh --agent opencode      # ~/.config/opencode/skills
./install.sh --agent goose         # ~/.config/goose/skills
./install.sh --agent all           # every agent already present on this machine
./install.sh --dir /some/other/skills-dir

For anything not in that list, check where your agent looks for skills and point --dir at it. The skill folders themselves are portable and need no changes.

Project-scoped install

To make the skills available to everyone working in one repo (and to Claude Code cloud sessions, which load skills committed to the cloned repository):

./install.sh --agent claude-project   # ./.claude/skills
./install.sh --agent codex-project    # ./.agents/skills

Then commit the result.

Uninstall

rm -rf ~/.claude/skills/whats-next ~/.claude/skills/human-context-rebuild
# or, if installed as a plugin:
claude plugin uninstall yskills@yskills
claude plugin marketplace remove yskills

The skills

whats-next

You run many sessions in parallel. You come back after half a day and can't remember where any of them stands — and every session greets you with a blank prompt.

Type /whats-next and the session re-orients you instead:

1. First line, always: can this session be closed right now? ("✅ safe to close — nothing running, everything on disk" / "⚠️ not yet — X is still running / Y only exists in this conversation"). If it's not closable, it names the single cheapest action that would make it closable.
2. A 3–5 line recap of what the session is about and where it stopped.
3. The next step as clickable choices (Yes/No, A/B/C/D) — concrete actions like "deploy to staging and run acceptance", never "continue the work". Pick one and it executes immediately. No blank prompt to fill in.

Rules baked into the prompt: only trust actual tool-call records (a step that was described but never executed didn't happen); never fabricate results for anything still running; never re-ask a decision that was already made.

It also works in a fresh context with zero history: point it at a worktree or branch (/whats-next in a new session, or "这个 worktree 做到哪了") and it rebuilds the state from git evidence instead — uncommitted diffs, branch commits vs. base, handoff/task docs — and the verdict becomes "can this worktree be wrapped up (merged/PR'd/deleted)?"

human-context-rebuild

The companion skill: a pure recap, no decisions. Type /human-context-rebuild (or just say "remind me what we're doing" / "我们到哪一步了") and get a sub-30-second re-orientation:

**现在在做 / Now**: …
**为什么 / Why**: …
**目标 / Goal**: …
**进展 / Done**: …
**下一步 / Next**: …

Under 150 words, anchored on file paths, branch names, and confirmed decisions — the things your brain latches onto faster than prose. Bilingual by design: it matches whatever language the session is running in.

Repo layout

yskills/
├── skills/                        # the skills themselves — portable, copy anywhere
│   ├── whats-next/SKILL.md
│   └── human-context-rebuild/SKILL.md
├── .claude-plugin/
│   ├── plugin.json                # Claude Code plugin manifest
│   └── marketplace.json           # marketplace catalog (this repo is its own marketplace)
├── install.sh                     # cross-agent installer
└── scripts/
    ├── validate.py                # Agent Skills spec validation
    └── package.sh                 # build dist/*.zip for claude.ai upload

Contributing

Add a skill as skills/<name>/SKILL.md, with frontmatter limited to the six fields the spec allows — name, description, license, compatibility, metadata, allowed-tools. name must match the directory name, and description must say both what the skill does and when it should trigger, in under 1024 characters. Anything beyond those fields is Claude Code-only and breaks claude.ai upload.

Then check it:

pip install pyyaml
python3 scripts/validate.py      # spec compliance
claude plugin validate .         # plugin + marketplace manifests

Both run in CI on every push and pull request.

License

MIT