Stop teaching Claude what it already knows.
Humans write for humans. They explain context, frameworks, concepts — because other humans need that scaffolding.
Claude doesn't.
delta strips everything Claude already knows from your inputs and processes only the delta — the new information, the specific problem, the unknown.
Debugging (full mode):
Before (41 tokens):
"I'm working on a Node.js Express API and I'm getting a 401
unauthorized error when I try to call the endpoint. I'm passing
the JWT token in the Authorization header."
After (12 tokens):
"401 on endpoint. JWT in Authorization header."
Code review (full mode):
Before (29 tokens):
"Could you please review this Python function and tell me if
there are any issues or improvements I could make?"
After (6 tokens):
"Review. Issues + improvements."
Explanation (ultra mode):
Before (19 tokens):
"I was wondering if you could explain how database
connection pooling works in simple terms?"
After (4 tokens):
"connection pooling. simple."
Same intent. Same answer quality. 40-80% fewer input tokens.
npx skills add shouvik12/deltaOr via Claude Code plugin marketplace:
claude plugin marketplace add shouvik12/delta
claude plugin install delta@deltaJust activate it:
/delta — activate (full mode default)
/delta lite — strip pleasantries only (~20% reduction)
/delta full — strip known context (~40-60% reduction)
/delta ultra — strip everything (~60-80% reduction)
stop delta — back to normal
Social noise — pleasantries, permission requests, hedging, filler
"Hey Claude, could you please help me..." → gone
"I was wondering if you might be able to..." → gone
"I think maybe possibly..." → gone
Known context — frameworks, concepts, stack introductions
"I have a React app using hooks..." → gone (Claude knows React)
"As you probably know, Python uses..." → gone
"My app uses Node, Express, MongoDB..." → gone (once is enough)
Negative space — flipped to positive
"don't be too long" → "brief"
"don't use bullets" → "prose"
"don't be technical" → "simple"
What delta never strips:
- The actual error, bug, or problem
- Numbers, file names, variable names
- Code blocks — never touched
- Anything Claude could NOT know without being told
| Level | Strips | Reduction |
|---|---|---|
lite |
Social noise only | ~20% |
full |
Social noise + known context | ~40-60% |
ultra |
Everything — bare signal only | ~60-80% |
You're wasting 40% of your tokens explaining things Claude already knows.
Caveman cut output tokens by making Claude talk like a caveman.
delta cuts input tokens by making YOU talk like Claude thinks.
They compose perfectly:
npx skills add JuliusBrussee/caveman # Claude talks less
npx skills add shouvik12/delta # You send less-
/delta:compress <file>— compress a file's prose before sending - Auto-detect conversation context and skip already-established facts
- Token count display — show savings per message
MIT
∆ if delta save you token, leave star ⭐