🎭 emoji-speak

Talk to Claude in 100% emoji. Fun first — and saves output tokens for verbose questions (38–63% in our examples).

A Claude Code skill / plugin that makes Claude reply in emoji-only prose while keeping all code, paths, URLs, and errors verbatim. Reduces output tokens for long-form questions, ships with a benchmark harness, and is the playful sibling of caveman — same "stop wasting tokens" idea, different vibe.

> Why does my useEffect run twice?
🤔 `React.StrictMode` → 🔁 dev-only double-invoke → 🎯 surface 🐛 in effects.
✅ prod: 1️⃣. ❌ not-bug.

Code, paths, URLs, error messages, and version numbers stay verbatim. Everything else becomes emoji.

🔀 Before / After

🗣️ Normal Claude · 115 tokens In JavaScript, the main difference between let and const is reassignment. Variables declared with let can be reassigned later in the same scope, while variables declared with const cannot be reassigned after their initial value is set. Both are block-scoped (only visible within the nearest enclosing { }), unlike var which is function-scoped. Note that const only prevents reassignment of the binding itself — if the value is an object or array, you can still mutate its contents. Use const by default for safety, and reach for let only when you genuinely need to reassign.	🎭 emoji-speak · 71 tokens (−38%) ≡ const ≠ let: let: ✏️ reassign ✅ const: ✏️ reassign ❌ (binding) — 🧭 mutate object/array ✅ both: block-scoped ({ }) ✅ const 1️⃣, let only when needed.
🗣️ Normal Claude · 110 tokens Your React component is re-rendering on every scroll event because the scroll handler is calling setState (or something that updates state) on each event. When state updates, React schedules a re-render. The fix is to either throttle/debounce the scroll handler so it fires less often, or move the scroll-tracked value out of state when you don't actually need to render with it. Wrap the handler in useCallback so its identity is stable, and consider passing the value via a ref instead of state if it's only used for layout calculations.	🎭 emoji-speak · 47 tokens (−57%) 🤔 scroll handler → setState every event → 🔁 rerender storm. 🔧 useCallback + throttle(16ms), or 📐 layout-only value via useRef (❌ state).
🗣️ Normal Claude · 49 tokens Yes — but make sure all tests pass first, then push to main. The CI pipeline should pick up the commit and trigger the deploy. Watch the CI dashboard while it runs. If anything fails, roll back via git revert and redeploy.	🎭 emoji-speak · 18 tokens (−63%) 🧪 ✅ ? → 🚀 git push origin main → 🏁

Same answer. Fewer tokens. Brain still big. Token counts from tiktoken o200k_base. Pairs are illustrative — your real savings depend on the question shape (verbose explanations compress well, one-line answers can get longer). Run the benchmark on your own questions.

---

⚡ Install (30 seconds)

In a Claude Code session:

/plugin marketplace add sytungan/emoji-speak
/plugin install emoji-speak@emoji-speak
/reload-plugins

---

🎮 Use it

To do this	Type this
Turn on	/emoji-speak:emoji on  or  "emoji on"  or  "use emoji"
Turn off	/emoji-speak:emoji off  or  "stop emoji"  or  "normal mode"

Your choice persists to the next session.

---

🌟 More demos

What does git rebase --onto do?

🌿 A → 🌿 B (base ⬅️) → 🌿 C ⤴️ → ✏️ replay commits onto 🌿 C. 🎯 surgical reparent.

Deploy to prod?

🧪 ✅ ? → 🚀 git push origin main → 🏁

drop the users table (safety auto-engages)

⚠️ Warning: this permanently deletes everything in the users table and cannot be undone. Verify you have a backup first.

DROP TABLE users;

🎭 resume. 💾 ✅?

---

🎬 Real outputs (verbatim from a live session)

Not curated examples — these are unedited replies Claude actually produced with emoji-speak active.

who are you?

👋 🤖 = Claude, 🏗️ Anthropic — claude-opus-4-7 1M 🪟. 🎭 🟢. ⏭️ ❓ ?

summary about history of apple til now

🗓️	🌟
1976-04-01	🏗️ in 🅿️ — Steve Jobs + Steve Wozniak + Ronald Wayne. Apple I 📜.
1984	🚀 Macintosh → 🖱️ GUI 👥. 📺 1984 ad 🔥.
1997	Jobs ⤴️ via NeXT 🛒. 🍎 ⏳ near 💀 → 🔄.
2001	🚀 iPod 🎵 + iTunes + 🥇 Apple Store 🏬.
2007-06	🚀 iPhone → 📱 era 💥.
2011-10-05	Steve Jobs 💀. Tim Cook 🪑 CEO.
2018-08	🥇 $1T 📊 cap.
2020-11	🚀 Apple Silicon M1 → 🧠 🍎-🏗️ → ⚡🔋.
2024-06	🚀 Apple Intelligence 🤖.

🧠 ⏮️ = 2026-01 → ⏳ post → ❓.

Notice: dates, proper nouns (Steve Jobs, Macintosh, iPhone), version-like names (M1), and currency ($1T) all stay verbatim. Everything else becomes emoji.

---

🛡️ When emoji steps aside

Plain English takes over for:

• 🚨 security warnings
• 🗑️ destructive / irreversible actions
• 1️⃣2️⃣3️⃣ multi-step procedures where wrong order = harm
• ❓ when you say "wait, what?" and need a clear answer

Once the dangerous part is over, emoji mode resumes automatically.

---

📊 Numbers

A benchmark harness lives in evals/. It runs each of 10 dev questions through three system-prompt conditions and counts output tokens with tiktoken o200k_base:

Arm	System prompt
baseline	none
terse	Answer concisely.
emoji_speak	Answer concisely. + the SKILL body

The script reports two deltas, both honest:

• emoji_speak vs baseline — what casual users will actually feel (most run Claude with no system prompt).
• emoji_speak vs terse — apples-to-apples; isolates the emoji style itself.

The honest framing: emoji-speak isn't uniformly a token saver. For questions where Claude's natural answer is already terse (like "how do I delete a remote git branch?"), wrapping a one-line answer in emoji + code spans actually adds tokens. For questions where Claude's natural answer is verbose (explanations, comparisons), emoji-speak compresses meaningfully — like the 33% saving on the let vs const example above. Run the harness on your own questions to see how it does for your workload:

cd evals && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python llm_run.py && python measure.py

---

🗺️ Roadmap

✅	Core SKILL + SessionStart hook
✅	/emoji-speak:emoji on|off
✅	Natural-language toggling + per-turn drift protection
✅	Benchmark harness
🚧	emoji-compress — convert your CLAUDE.md to emoji
💭	Statusline indicator
💭	Multi-agent (Cursor, Windsurf, Gemini, Codex, Cline, Copilot)

---

🧪 Develop

node --test tests/*.mjs    # run all 25 tests

(Node 24+ needs the explicit glob — tests/ alone doesn't recurse.)

---

❓ FAQ

How do I save tokens with Claude Code? Install emoji-speak. For verbose questions (explanations, comparisons, debugging), it cuts 30–60% of output tokens by replacing prose with emoji. Code, paths, and errors are preserved exactly.

Does emoji-speak work with the Claude API? Currently it's packaged as a Claude Code plugin. The SKILL itself (skills/emoji-speak/SKILL.md) is a system-prompt fragment — you can copy its body into any Anthropic SDK call's system field and get the same behavior.

Will it break my code output? No. Anything inside `backticks`, fenced code blocks, file paths, URLs, error messages, version numbers, and proper nouns is preserved verbatim. Only English prose becomes emoji.

Is it just a gimmick? The fun is the point. The token savings are a bonus — measured per-question in evals/. For one-line answers (e.g. "how do I delete a remote git branch?") emoji-speak can actually use more tokens because the wrapping costs more than the saved word. For verbose answers it wins consistently.

How is this different from "Answer concisely."? The benchmark compares both. Plain "Answer concisely." does shorten output, but emoji-speak shortens it further on long-form questions while staying readable as a style, not just a length cap.

Does it support Cursor / Windsurf / Gemini / Codex? Not yet — Claude Code only for now. Multi-agent manifests are on the roadmap (same approach as caveman).

---

⭐ Star This Repo

If emoji-speak makes you smile (or saves you tokens), drop a star. Helps other people find it. 🎭

---

📜 Credits

• Inspired by caveman — same plugin scaffolding, opposite vibe.
• MIT licensed. See LICENSE.