Laconic 🪙

from Laconia (Sparta). Spartan speech style: maximum meaning, minimum words.

Claude Code plugin. Spartan brevity. Inspired by 🪨 caveman — same goal, measured savings (see Benchmarks).

How it works

• SessionStart hook (bash) injects the rule once per session — laconic mode auto-active.
• /laconic skill reinforces caps if needed: ≤20 words or ≤3 lines, no preamble, no headers, no trailing summaries.
• Exit with stop laconic or normal.

Install

One-liner

curl -fsSL https://raw.githubusercontent.com/brillcp/laconic/main/install.sh | bash

Adds the marketplace and enables the plugin in ~/.claude/settings.json. Restart Claude Code — laconic auto-activates every session via the SessionStart hook. No /laconic needed.

Manual

/plugin marketplace add brillcp/laconic
/plugin install laconic@laconic

Then /reload-plugins (or restart Claude Code).

Uninstall

curl -fsSL https://raw.githubusercontent.com/brillcp/laconic/main/uninstall.sh | bash

Removes the marketplace + disables the plugin in ~/.claude/settings.json. Backup written as settings.json.laconic.bak.

Use

• /laconic <prompt> — reinforce mode mid-session
• /laconic:stats — estimate tokens saved this session
• /laconic:audit — list enabled plugins, flag unused ones, prompt before disabling
• /laconic:benchmark — run the benchmark from anywhere and print the full results table in the CLI (pass RUNS=N for more samples)

Statusline indicator

Optional. Add to ~/.claude/settings.json:

"statusLine": {
  "type": "command",
  "command": "~/.claude/plugins/cache/laconic/laconic/<sha>/statusline/statusline.sh"
}

Shows 🪙 LACONIC · ~N tokens saved live in the status bar — counter reads the session transcript and applies a 1.74 expansion factor to estimate savings vs normal mode. The factor varies by model — Opus benefits more, Haiku less — so re-run the benchmark with higher RUNS to refresh it for your setup.

Further token-saving tips

Laconic trims output tokens. To trim more:

• Match the model to the task. Opus output costs ~15× Haiku per token, so each saved token saves more dollars on Opus — laconic + Opus = the biggest absolute savings if you need Opus's reasoning. For trivial work (renames, doc tweaks), /model haiku is plenty. Set the default in ~/.claude/settings.json ("model": "claude-opus-4-7", etc.) and switch mid-session with /model.
• Audit plugins with /laconic:audit — every enabled plugin loads tool/skill metadata each turn.
• Trim verbose CLAUDE.md files, skill description: fields, and hook output — they ride every request.
• Read narrowly (built-in): laconic's rule already steers Claude to Grep -n pattern file first, then Read offset:N limit:M around the match — instead of reading whole files. Sub-agents recommended for noisy searches so their context stays isolated.
• Push back, don't comply blindly (built-in): the rule weighs performance and best practices and pushes back on bad designs tersely, so brevity doesn't trade away architectural judgment.

Update

/plugin update laconic

Benchmarks

Output tokens for 10 language-agnostic prompts. Claude Code, Haiku 4.5, RUNS=1 (single run — re-run with higher RUNS for median).

#	Prompt	Off	On	Saved
1	Mutex vs semaphore	973	691	28%
2	TCP vs UDP	1141	399	65%
3	Binary search time complexity	864	661	23%
4	JWT authentication end to end	1387	715	48%
5	REST vs GraphQL tradeoffs	784	775	1%
6	CAP theorem	489	575	-17%
7	Processes vs threads	317	1204	-279%
8	HTTPS handshake	392	1099	-180%
9	Big O / Theta / Omega	1525	556	63%
10	Dependency injection	1843	759	58%
	Total	9715	7434	23%

Note: previous totals were single-run and showed variance (e.g. prompt 6 spiked to -29%). The benchmark now:

• runs each prompt RUNS times per mode (default 3) and reports the median, which damps single-run noise.
• isolates the rule's effect with LACONIC_BENCH=1 (skips the SessionStart hook for the Off runs) and --append-system-prompt for the On runs — apples-to-apples, no live-settings mutation.

Expected savings depend on the model: Opus is more verbose by default, so it has more to trim (~40%); Haiku is already concise (~20-25%). Re-run with your model to refresh results.md.

Reproduce

git clone https://github.com/brillcp/laconic && cd laconic
brew install jq            # macOS — apt/dnf elsewhere
RUNS=5 ./benchmark/run.sh  # 5 samples per prompt per mode → median

Edit benchmark/prompts.txt to swap in your own prompts. Output → benchmark/results.md. Default RUNS=3 if unset.