Terse — Token-Efficient Communication for Claude Code
Built on the brilliant Caveman Plugin — adopts caveman's smart-caveman voice for stickiest compression, adds a deliverable-vs-explanatory code split so production code stays untouched.
A Claude Code plugin that compresses prose, eliminates filler, and reduces explanatory code to minimal form — without touching deliverable code quality. Pre-caveman benchmarks measured ~36% median word reduction; the caveman-voiced rewrite targets ~70%, pending re-benchmarking.
What It Does
Terse mode transforms Claude's communication style:
| Normal | Terse |
|---|
| "Sure! I'd be happy to help you with that. The issue you're experiencing is likely caused by a misconfigured authentication middleware..." | "Bug in auth middleware. Token expiry check use < not <=. Fix:" |
| 6-line PostgreSQL pool setup with imports | `new Pool({ max: 20, idleTimeoutMillis: 30000 })` + one-line explanation |
Key Behaviors
- Prose: Drops articles, filler words, pleasantries, and hedging. Fragments are fine.
- Explanatory code: Compressed to one-liners, pseudocode, or name-drops. Full snippets only for obscure APIs.
- Deliverable code: Untouched. Production code stays production-quality.
- Technical terms: Always exact. No simplification of domain language.
Install
From marketplace
claude plugin install terse@fingerskier-plugins
Manual install
git clone https://github.com/fingerskier/terse-claude-plugin.git ~/.claude/skills/terse
# Or copy directly
cp -r terse-claude-plugin/skills/terse ~/.claude/skills/
Usage
Activate in any Claude Code conversation:
/terse
Or say start terse, terse mode, caveman mode, less tokens, be brief.
Deactivate:
stop terse
Who Is This For?
Experienced developers who don't need hand-holding.
If you already know what useMemo does and just need to know when to reach for it, terse mode skips the tutorial and gives you the answer.
Token Savings Estimate
| Response Type | Estimated Savings |
|---|
| Pure explanation | 35–55% |
| Mixed code + explanation | 25–50% |
| Short direct questions | 40–60% |
| Mostly deliverable code | 0–10% |
| Median across categories | ~36% |
Savings vary by prompt and conversation. Numbers based on automated benchmarking of 26 prompts across 8 categories.
Test Results
We ran 26 prompts across 8 categories, comparing normal vs terse output by word, character, and line count.
Some normal-mode runs returned empty responses (test harness issue), so results below include only valid pairs where both modes produced content.
| Category | Valid / Total | Word Reduction |
|---|
| Pure explanation | 3 / 4 | 12–58% |
| Mixed code + explanation | 3 / 4 | 26–52% |
| Error & technical terms | 3 / 3 | 20–47% |
| Short direct questions | 4 / 4 | -3–25% |
| Multi-step tutorial | 2 / 3 | 6–16% |
| Pleasantry traps | 3 / 3 | -74–25% |
| Boundary cases (should not compress) | 1 / 2 | -299–26% |
Takeaways
- Best gains on prose-heavy responses. Explanations and mixed code/explanation see 20–58% word reduction consistently.
- Diminishing returns on short answers. Already-concise responses (short questions, error diagnostics) see 0–25% reduction.
- Inconsistent on pleasantries and boundary cases. Terse sometimes expanded output — particularly on prompts asking for structured deliverables like PR descriptions, where terse mode added more detail than normal mode provided.
- Test harness gaps. 5 of 26 normal-mode responses came back empty, inflating negative percentages in the raw summary. Valid-pair results are more representative.
Full results are in test/results/summary.txt.
License
MIT
fingerskier
