Cut your AI costs by up to 72% — using Singapore English grammar.
AI assistants (ChatGPT, Claude, Copilot) charge based on how many words they produce. These words are called tokens — and fewer tokens means lower cost.
SingCompress tells the AI to respond using Singlish — a real language spoken by millions in Singapore — which naturally says the same thing in fewer words. Think of it as shorthand for AI, but based on a real spoken language with decades of linguistic research.
Inspired by caveman, but linguistically grounded.
All examples from Claude Opus 4.6:
| Default | Siudai | Medium | Gao | |
|---|---|---|---|---|
| SQL Injection Review | "This code has a critical security flaw: it's vulnerable to SQL injection. String interpolation inserts user input directly into the SQL query with no sanitization..." | "SQL injection vulnerability — attacker can pass '; DROP TABLE users; -- as username and destroy your data. Fix: use parameterized queries." |
"SQL injection vulnerability sia. User pass '; DROP TABLE users;-- as username, GG liao. Use parameterized query." |
"SQL injection sia. Jialat. Use parameterized q lor. Raw f-string → anyone can inject one." |
| CORS Error | "This is one of the most common development issues. The browser blocks requests between different origins (different ports count as different origins). Here are the solutions..." | "Install cors middleware on your Express server. For production, lock down to specific origin." |
"Classic CORS issue. Frontend and backend diff ports → browser block cross-origin req. Fix: Install and use cors middleware in Express." |
"Classic one. Install cors mw. Want restrict to your React port only, use cors({ origin: 'http://localhost:3000' })." |
| Stale Closure | "This is one of the most common React pitfalls: stale closures. When your useEffect callback is created, it captures the state variable's value at the time of that render..." | "Classic React stale closure problem. Your useEffect callback captures state value from render it was created in, and if dependency array not updated properly, it keeps referencing old value." | "Stale closure lah. useEffect callback capture state value from render it was created. You update state → new render → but old effect still holding old value." | "Stale closure lah. React useEffect captures vals from render it was created in. If dependency array bo include that state var, cb still sees old value one." |
| React Re-renders | "The core issue is referential identity. Every time the parent renders, your inline object prop creates a new object in memory, even if its contents are identical..." | "Every time parent renders, inline object {} creates new reference in memory. React does shallow comparison for props — different reference means different prop, so child re-renders." |
"Inline obj → new ref every render → React sees different prop → re-render. Same thing for inline fn — use useCallback." |
"Inline obj = new ref every render → === check fail → child re-render cfm. Same thing for inline fns & arrays hor." |
| Race Condition | "This is a classic lost update / read-modify-write race condition. Your code likely does a read-then-write pattern concurrently..." | "Multiple requests doing read-then-write pattern concurrently. Both get same value, both write same increment, lost one update." | "Classic race condition lah. Got lost update problem. Two req read counter same time → both get same value → both write 6 → lost one increment." | "Cfm race condition one. Classic lost update problem sia. 2 req read counter = 5 same time → both write 6 → lost 1 increment. GG." |
| Security Headers | "There are several security issues and missing best practices in this setup: Wildcard CORS Origin is critical, setting origin: '*' allows any website to make requests to your API..." |
"Few security issues here: CORS wide open, no rate limiting, app vulnerable to brute force and DDoS..." | "Got few security issues here: CORS wide open — origin: '*' means any domain can hit your API. Kena CSRF attack easy." |
"Few issues sia: CORS origin: '*' — jialat, any domain can hit your API. Bo rate limiting — kena DDoS cfm GG. Bo helmet — app.use(helmet()) steady one." |
Same technical content. Up to 88% fewer tokens. Quotes are excerpted for brevity.
- Save money — Up to 72% fewer output tokens = proportionally lower API bills
- Faster responses — Fewer tokens to generate = quicker replies
- Linguistically grounded — Not random abbreviation. Based on Singlish, a rule-governed contact language with extensive linguistic research
- Quick Start
- How It Works
- Compression Levels
- More Examples
- Benchmarks
- Singlish Grammar Patterns
- Model Compatibility
- Running the Eval
- Why Singlish?
claude plugin marketplace add dljw/singcompress
claude plugin install singcompress@singcompressAuto-activates on every session. Use /singcompress siudai|medium|gao|off to switch levels.
npx skills add dljw/singcompress -a cursor
npx skills add dljw/singcompress -a github-copilot
npx skills add dljw/singcompress -a windsurf
npx skills add dljw/singcompress -a clineUses the Vercel Skills CLI to install the SKILL.md into your agent's skills directory.
Manual install (any agent)
Copy skills/singcompress/SKILL.md into your agent's skills directory:
# Claude Code
cp -r skills/singcompress ~/.claude/skills/singcompress
# Cursor
cp -r skills/singcompress .cursor/skills/singcompress
# Or just read the SKILL.md and paste it into your agent's system promptSingCompress is a system prompt that instructs the LLM to respond using Singlish grammatical patterns that naturally compress meaning. It is not a post-processor or proxy — the model generates compressed output directly.
What gets compressed: Natural language prose — explanations, descriptions, instructions.
What stays untouched: Code blocks, error messages, technical terms, file paths, URLs, JSON, YAML, SQL, shell commands.
Safety: Auto-Clarity mode automatically reverts to Standard English for security warnings, irreversible operations, and anything where fragmented language risks misinterpretation.
Drop articles, copulas, filler, pleasantries. Grammar stays intact.
"Server returning 500 error because database connection pool exhausted."
Full Singlish grammar: "got" existential, "liao"/"already" completion, PRO-drop, topic-comment, "can meh?", abbreviations (DB, auth, fn, etc.), arrow causality, aggressive fragmentation.
"SQL injection vulnerability lah. Never sanitise input → attacker can break out of string."
All medium rules plus SG particles (lor, sia, hor, one, leh), Singlish vocabulary (jialat, shiok, simi, sian, steady, bo), SG text shortcuts (cfm, tmr, alr, nvm), symbol compression (&, w/, b/c).
"SQL injection sia. Jialat. Use parameterized q lor — DB driver handle escaping, safe liao."
See more before/after examples (Sonnet)
Default (1032 tokens): "This is one of the most common issues in local full-stack development! Here's a thorough breakdown of what's happening and how to fix it..."
Medium (162 tokens): "Install
corspkg, then add to Express:app.use(cors({ origin: 'http://localhost:3000' }));Put before all routes lah."
Default (870 tokens): "This is a very common React gotcha known as a stale closure. Here's what's happening and why..."
Medium (179 tokens): "Classic stale closure lah. useEffect captures state value at render time. Got deps array wrong (or empty
[]), so callback never re-created → always sees old value."
Default (1587 tokens): "What you're experiencing is a classic race condition caused by a non-atomic read-modify-write cycle..."
Heavy (431 tokens): "Race condition / lost update lor. Multiple reqs read same val → each increment own copy → write back → earlier write overwritten."
Medium mode achieves 72% token reduction on Sonnet and beats generic "be concise" prompts by 16-32% across all models.
Full benchmark results — tested across 24 prompts on 3 Claude models
| Level | Sonnet | Opus | Haiku |
|---|---|---|---|
| Light (siudai) | -64% | -48% | -33% |
| Medium | -72% | -60% | -44% |
| Heavy (gau) | -65% | -61% | -28% |
| Level | Sonnet | Opus | Haiku |
|---|---|---|---|
| Light | -12% | +9% | -4% |
| Medium | -32% | -16% | -19% |
| Heavy | -16% | -19% | +5% |
Medium mode beats a generic concise prompt by 16-32% across all models — Singlish compression is structurally more efficient than just "be brief".
| Category | Light | Medium | Heavy |
|---|---|---|---|
| Coding explanation | -62% | -69% | -62% |
| Debugging | -66% | -76% | -70% |
| Architecture | -64% | -72% | -66% |
| Code review | -63% | -74% | -71% |
| Multi-step | -59% | -67% | -58% |
| Conversational | -70% | -75% | -65% |
Full table of Singlish grammar patterns used
| Pattern | Standard English | SingCompress | Tokens Saved |
|---|---|---|---|
| "Got" existential | "There is a bug in line 42" | "Got bug in line 42" | 2-3 |
| "Already" / "liao" completion | "I have already fixed it" | "Fix already" / "Fix liao" | 3-4 |
| PRO-drop | "I cannot find the file" | "Cannot find file" | 2 |
| Topic-comment | "This is the function that handles auth" | "This function — handle auth" | 4-5 |
| Bare conditionals | "If you get an error, restart" | "Got error, restart" | 3 |
| "Can/Cannot" | "Yes, that's possible" | "Can." | 3-4 |
| "Lah" emphasis | "No, that definitely won't work" | "Cannot lah" | 4-5 |
| "Can meh?" doubt | "Are you sure that will work?" | "Can meh?" | 5-6 |
| Arrow causality | "This causes that which leads to" | "X → Y → Z" | 3-5 |
| Copula omission | "The server is busy" | "Server busy" | 2 |
| "Jialat" (heavy) | "This is seriously broken" | "Jialat" | 3-4 |
| "Sia" emphasis (heavy) | "This is surprisingly fast" | "Fast sia" | 3-4 |
| "Lor" resigned (heavy) | "Well, that's just how it is" | "Like that lor" | 6-7 |
| SG text shortcuts (heavy) | "confirm", "tomorrow", "already" | "cfm", "tmr", "alr" | 1-2 each |
| Symbol compression (heavy) | "with", "without", "because" | "w/", "w/o", "b/c" | 1-2 each |
These are not random simplifications — they are documented features of Singlish grammar, a rule-governed contact language with extensive linguistic research.
| Model | Singlish Quality | Best Level | vs Default | vs Concise |
|---|---|---|---|---|
| Claude Sonnet | Fluent, natural | Medium | -72% | -32% |
| Claude Opus | Precise, idiomatic | Heavy | -61% | -19% |
| Claude Haiku | Functional | Medium | -44% | -19% |
All three models beat generic "be concise" prompts. Sonnet delivers the highest compression. Opus is the only model where heavy outperforms medium — it fully leverages the expanded Singlish vocabulary. Haiku works well on medium but struggles with the heavy rule set.
Eval instructions (for developers)
# Install dependencies
npm install
# Run full eval (uses your Claude Code subscription)
npx tsx evals/run.ts --model sonnet
# Run a single prompt
npx tsx evals/run.ts --prompt-id debug-cors --model sonnet
# Analyze results
npx tsx evals/measure.ts
# With semantic preservation scoring
npx tsx evals/measure.ts --with-semanticSinglish evolved as a contact language among multilingual speakers in Singapore. Its grammar systematically drops redundant elements — articles, copulas, subject pronouns, tense markers — while adding efficient particles ("lah", "liao", "got") that pack pragmatic meaning into single tokens. These aren't deficiencies; they're compressions that evolved to communicate meaning quickly.
SingCompress applies these same compressions to LLM output. The result is fewer tokens with the same technical content — and if you're Singlish-literate, it reads naturally.
MIT