Dense Mode

Reduce Claude's output tokens by ~59% without losing information or readability.

Dense mode compresses at the structural level — arrow notation replaces sentences, key:value replaces paragraphs — instead of just dropping articles like cave talk. The result reads like a senior engineer's Slack message, not broken English.

Install

npx skills add 0xpili/dense-mode

Usage

/dense          # Default — arrows, key:value, no filler
/dense lite     # Keep grammar, just cut filler and hedging
/dense ultra    # Maximum compression, telegraphic style
/dense off      # Back to normal

Example

Before (328 tokens):

This is a classic clock skew issue. The exp (expiration) claim in your JWT is being compared against the server's current time, and there's likely a mismatch between the clock on the machine that issued the token and the machine that's validating it. Here's what's probably happening and how to fix it...

After (124 tokens):

Cause: clock skew between token issuer and validator. Server clock slightly ahead → fresh tokens appear expired.

Fix: jwt.verify(token, secret, { clockTolerance: 30 });

Also check: exp in seconds not ms, timezone mismatch, cross-env tokens.

Benchmarks

Manual comparison

Claude wrote the same 6 responses in each style, measured via word count × 1.3 (full research):

Metric	Normal	Cave Talk	Dense Mode
Output reduction	—	54%	59%
Info completeness	5.0	4.8	5.0
Readability	5.0	4.3	4.5
30-turn savings	—	80K tokens	88K tokens

Dense mode beats cave talk by 5 percentage points with zero information loss.

Automated CLI benchmark

Real claude -p runs with --output-format json measuring actual API token counts. 6 prompts × 3 configs × 3 trials (median), Sonnet model. See benchmark/ for raw data and script.

Prompt	Type	Baseline	Caveman	Dense
React re-render	prose	406	472	527
JWT middleware fix	code	1,319	2,341	1,196
Microservices vs mono	prose	610	523	392
Security review	mixed	894	805	1,003
Docker multi-stage	code	1,505	1,684	2,224
Refactor + tests	tool	1,053	962	923
Average		964	1,131	1,044

Why it works

Cave talk compresses at the word level — dropping articles (a, the, an), using fragments. Dense mode compresses at the structural level:

• A → B → C replaces "A causes B, which leads to C"
• Status: active, Scope: auth replaces "The status is currently active and the scope covers the auth module"
• One sentence per concept replaces multi-sentence explanations

Structure-level compression is more efficient because it removes redundant grammatical scaffolding, not just small words.

What it preserves

• Code blocks (unchanged)
• Technical terms (exact)
• Error messages (verbatim)
• Security warnings (full reasoning)
• File paths (complete)

Research

The research/ folder contains the manual empirical comparison:

• dense-vs-cave-talk.md — Full write-up with 6 findings
• data/results.md — Raw comparison tables and composite scorecard
• data/test-cases.md — All 6 prompts with side-by-side responses in 4 styles

The benchmark/ folder contains the automated CLI benchmark:

• run-benchmark.sh — Reproducible script using claude -p --output-format json
• prompts.json — 6 test prompts
• results/ — Raw JSON output and summaries from all runs

Part of Where the Tokens Go — research on Claude Code token optimization from the inside.

License

MIT