Claude Code Skills

Skills I built for my own Claude Code workflow and kept reaching for. Sharing them here in case they help you too.

/plugin marketplace add enkr1/claude-code-skills
/plugin install enkr-skills@enkr-skills

Skills

Skill	What it does
threads	A branching task tracker that never drops a ball: per-chat trees, an all-chats view, auto-reinjected so Claude never forgets.
commit	Gitmoji and conventional commits, with anti-AI-signature rules.
comprehensive-review	Staff-engineer ship gate: reviews the design, composes the built-in /code-review for the diff, then gives one decisive verdict.
md2pdf	Markdown to clean, print-ready PDF via pandoc and headless Chromium (no LaTeX). Cross-browser, batch, with em-dash and page-count checks.

---

✦ threads: never lose a task thread again

A branching task tracker that lives inside Claude Code, built for ADHD and nonlinear work. You dump branching ideas faster than you can file them; threads captures each one into a tree, reminds Claude of your open threads on every prompt so it stops forgetting, and never drops a ball.

here is your default glance, just where you are and one layer around it:

  threads · web-app                         6 open · 0 parked
  ───────────────────────────────────────────────────────────
  … ship v2 › payments
  ▸ ● refund flow
     ├─ ○ proration math
     └─ ○ webhook retry  (+3 deeper)
  ───────────────────────────────────────────────────────────

all zooms out to every chat's full tree, grouped by project:

  threads · all                             2 chats · 5 open
  ───────────────────────────────────────────────────────────
  ◆ dashboard · ship v2
  ○ ship v2
  ├─ ○ write tests
▸ └─ ● fix auth redirect
  ◆ side-project · launch
  ○ launch
▸ └─ ● wire up payments
  ───────────────────────────────────────────────────────────

• Per-chat trees: each conversation is its own tree with its own "you are here" (▸). Chats never tangle.
• Auto-reinjected: a UserPromptSubmit hook feeds your open threads back into context every turn. This is the part a plain markdown skill cannot do, and it is why Claude stops forgetting parked work.
• No dropped balls: a task leaves the tree only when it is done or you kill it. Never silently. Stale ones are flagged, not deleted.
• Self-pruning: finishing a parent (done) shifts its children up to the grandparent and drops the parent, so the tree stays shallow and shows only what is still open. compact runs that sweep across a whole backlog at once; done leaves stay as ✓.

Glyphs: ▸ here, ● active, ○ paused, ✓ done, ✕ blocked, ◦ snoozed.

Use it by talking

"switch to X", "bt" / "backtrack", "done", "snooze X", "where am i", or just dumping new tasks all drive it. Or call the CLI directly:

threads  capture "<name>" | switch "<q>" | bt | done [q] | snooze "<q>" [days] | compact | here (default) | all

Install threads

/plugin marketplace add enkr1/claude-code-skills
/plugin install enkr-skills@enkr-skills
node ~/.claude/skills/threads/install.mjs   # wires the hook, then restart Claude Code

Why the extra step: plugin-defined hooks cannot inject context yet (claude-code#12151), so install.mjs adds the UserPromptSubmit hook to your settings.json (backed up, append-only, idempotent). To uninstall, delete that one hook block.

Statusline (optional)

threads ships a one-line "where am i" for the Claude Code statusline (statusline/threads-statusline.mjs). To show it, add this to your statusline script, passing the chat's session_id:

SID=$(jq -r '.session_id // empty')
TH=$(THREADS_SESSION="$SID" node ~/.claude/skills/threads/statusline/threads-statusline.mjs 2>/dev/null)
if [ -n "$TH" ] && [ "$TH" != "threads: idle" ]; then printf "\n%s" "$TH"; fi
exit 0

The exit 0 is not optional. If a statusline script exits non-zero, Claude Code blanks the entire bar. A trailing && chain that short-circuits when there is nothing to show is an easy way to exit 1 by accident, so always end the script with exit 0. If node is not on the statusline's PATH (for example under nvm), use the absolute node path.

Local-first: state is a single JSON file at ~/.claude/threads.json. No network, no account.

---

✦ commit: gitmoji + conventional commits

Skill	Description
commit	Gitmoji + conventional commits with anti-AI-signature rules

Enforces a clean, single-line format <emoji> <type>(<scope>): <description>, with a "Must Look Human" section that strips Co-Authored-By and "Generated with Claude Code" so commits read like you wrote them.

✨ feat(auth): add OAuth2 PKCE flow for mobile clients
🐛 fix(api): prevent race condition on concurrent profile updates (#163)

---

✦ comprehensive-review: a staff-engineer ship gate

Not another code reviewer, it composes the built-in /code-review and adds what it lacks. Run it before you ship:

1. Design review (when no code exists yet), critique the approach before you build. /code-review cannot do this.
2. Code review, delegates the diff to /code-review (its effort tiers, cloud review, --fix).
3. Self-verify, every finding must survive being argued against, so it does not cry wolf.
4. Verdict, one decision: REJECT / NEEDS WORK / APPROVE, plus what is blocking ship.

For a raw diff review, call /code-review directly. Use this when you want design judgment plus a go/no-go.

---

✦ md2pdf: Markdown → print-ready PDF

No LaTeX, no fiddling. One command turns a .md into a clean PDF, using pandoc plus a headless Chromium renderer (the same engine your browser prints with).

md2pdf.sh answers.md submission.pdf   # explicit output
md2pdf.sh notes.md                    # → notes.pdf next to it
md2pdf.sh a.md b.md c.md              # batch, each → its own .pdf

Built for graded submissions: it warns on em dashes (the AI tell), reports the page count and flags going over 2 pages, and auto-detects whatever Chromium-family browser you have (CHROME_BIN to force one). Restyle every PDF by editing a single scripts/md2pdf.css.

Contributing

Skills that solve real problems are welcome.

1. Fork this repo
2. Create skills/<your-skill-name>/SKILL.md (Agent Skills Spec)
3. Open a PR describing the problem it solves, example usage, and why it beats doing it by hand

Quality bar: opinionated (take a stance), compact (under 500 words; reference docs in a references/ subdir), no secrets, and actually tested across sessions.

Prior art and inspiration

threads is far from the first to think about this. With thanks to the work it learned from or shares ideas with:

• Discovery Trees articulate the ADHD branching-task pattern (branch off a tangent, return to the parent when it is done). That is the mental model threads automates.
• Beads by Steve Yegge treats a dependency graph as agent memory, the bar for durable, agent-owned task state.
• Task Master and Claude Code PM are prior art for task management inside Claude Code.
• Claude Code's hooks (UserPromptSubmit context injection) are what make the every-turn reinjection possible.

Missed an attribution? Open an issue and I will add it.

License

MIT