Terse Mode — Output Token Savings

Terse Mode is Kodelyth ECC's built-in output-token compressor. It changes how the AI talks — dropping filler while preserving every technical detail — for 40-70% output token savings on typical replies.

Inspired by Caveman (MIT, by Julius Brussee). ECC's implementation is independent: our own prompt, our own compressor, our own ledger, our own dashboard tile. No Caveman code copied, no Caveman dependency, no shell-out to Caveman binaries. Credit in every artifact.

The rule

Make the AI's mouth smaller, not its brain smaller.

Same answers. Fewer words. Code byte-exact.

Four dial levels

Type /terse [level] in your AI tool. The level sticks for the session.

Example — same fix, three levels:

Normal: The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object.

Lite: The React component re-renders because a new object reference is created on each render. Wrap the object in useMemo.

Full: New ref each render → re-render. Wrap object in useMemo.

Ultra: Ref/render. useMemo it.

What Terse mode NEVER touches

Byte-preserved in every level:

• Fenced code blocks ```lang ... ```
• Inline code
• Shell commands and error text
• URLs
• File paths, function names, identifiers
• Numbers, versions, hashes
• YAML / JSON / config snippets

Install

Terse mode is installed automatically when you install ECC:

npm i -g kodelyth-ecc
kodelythecc --target claude-code

The post-install step copies:

• ~/.claude/skills/terse-mode/SKILL.md — the prompt-level compressor rule
• ~/.claude/commands/terse.md — the /terse slash command
• ~/.claude/commands/terse-compress.md — the /terse-compress slash command

The skill stays dormant until you type /terse. This is deliberate — Terse mode changes how the AI talks, so it must be user-consented per session.

CLI reference

kodelythecc terse status                       # shipped / installed / ledger paths
kodelythecc terse stats [--json]               # turns tracked, tokens saved, level breakdown
kodelythecc terse compress <file> [--dry-run]  # rewrite a markdown file, byte-preserves code
kodelythecc terse enable [--target X | --all]  # install skill + slash commands into an IDE
kodelythecc terse --help                       # focused help

Slash commands (inside your AI tool)

/terse             # set to full (default)
/terse lite        # light trim
/terse full        # telegram-style fragments
/terse ultra       # maximum compression
/terse off         # restore normal voice

/terse-compress CLAUDE.md              # compress a file
/terse-compress tasks/lessons.md
/terse-compress ~/.claude/CLAUDE.md

Compress memory files — permanent savings

/terse-compress and its CLI counterpart kodelythecc terse compress <file> run a deterministic zero-dependency compressor on markdown files. This is different from the runtime /terse skill:

• The skill compresses runtime replies (session-scoped)
• The compressor compresses files on disk (permanent)

Perfect target: CLAUDE.md, tasks/lessons.md, rule files, and any memory context that gets injected into every session.

What the compressor does

• Strips 40+ filler patterns (in order to, due to the fact that, basically, essentially, ...)
• Converts wordy connectives to compact ones (in order to → to, due to the fact that → because)
• Merges wrapped-prose paragraphs, collapses whitespace
• Removes politeness padding ("You should", "I would recommend", "Please note")

What it byte-preserves

• Fenced code blocks
• YAML frontmatter (whole block between --- markers)
• Inline code
• URLs (https?://...)
• File paths (Unix, ~/, ./, ../, absolute)
• Markdown link URLs (inside [text](url))

Real numbers

On a real prose-heavy markdown file: ~30% byte reduction, 100% code/URL/path integrity, verified on real tests before every release. Deterministic — same input, same output, safe to re-run.

Programmatic usage

const { compressText, compressFile } = require('kodelyth-ecc/scripts/terse/compress.js');
 
// String-in, string-out
const { output, stats } = compressText(source);
console.log(stats);
// → { originalBytes, newBytes, saved, savedPct, estimatedTokensSaved }
 
// File in place, with backup
const result = compressFile('/path/to/CLAUDE.md', {
  write: true,
  backup: true,  // keeps original at CLAUDE.md.pre-terse.bak
});

Runtime savings ledger

Every terse-active turn is recorded in ~/.kodelythecc/terse/ledger.jsonl:

{
  "ts": "2026-07-04T17:29:33.851Z",
  "level": "full",
  "rawEstimate": 250,
  "actual": 125,
  "saved": 125,
  "source": "claude-code"
}

Query it:

kodelythecc terse stats
kodelythecc terse stats --json

Or view in the dashboard: kodelythecc dashboard → Token Savings tab → Output savings (Terse mode) section.

Combined with RTK

• RTK compresses input tokens (tool output → LLM)
• Terse mode compresses output tokens (LLM → user)

They're orthogonal. Stack together for 55-65% total token savings on typical coding sessions, 65-70% on explain-heavy or code-review sessions.

Baked into agents (Phase C)

Two existing ECC agents have opt-in terse mode built in:

• code-reviewer — when /terse is active, PR comments become one-line: L42: 🔴 bug: user null. Add guard.
• release-captain — when /terse is active, Conventional Commit subjects stay ≤50 chars, changelog entries stay compact

Attribution

• License: ECC's terse-mode implementation is MIT
• Design inspiration: Caveman by Julius Brussee (MIT)
• What we borrowed: The core insight ("compress the mouth, not the brain") and the level-dial pattern
• What we didn't borrow: Zero copied code. Different prompt, different levels (no wenyan Chinese mode), different compressor implementation, integrated with our RTK ledger for combined input+output tracking

See also

• RTK — the input-side companion
• Dashboard — live combined savings view
• Intent Routing v2 — how routing adapts announcement style when terse is active