gist

Adaptive verbosity for Claude Code. A heuristic classifier picks a level for every turn — minimal, terse, or full — and a code-enforced safety floor forces full prose on destructive, security-sensitive, or multi-step requests. Code, errors, and your language are always preserved verbatim.

How it works

• SessionStart hook injects a one-time rulebook describing the three levels.
• UserPromptSubmit hook classifies each prompt and injects THIS TURN: <level>.
• Statusline shows the current level and estimated tokens saved.
• /gist-stats reports measured output tokens, a per-level histogram, and a conservative savings estimate (full-prose turns count as zero savings).

Install

Add the marketplace and install the plugin:

/plugin marketplace add <path-or-repo>
/plugin install gist@gist-marketplace

Statusline (manual, one-time)

Claude Code plugins cannot set the main statusline automatically, so add this to your ~/.claude/settings.json (adjust the path to where the plugin is installed):

{
  "statusLine": {
    "type": "command",
    "command": "bash \"$HOME/.claude/plugins/marketplaces/gist-marketplace/gist/hooks/gist-statusline.sh\""
  }
}

Opt out of the badge any time with GIST_STATUSLINE=0 in your environment.

Levels

Level	When	Behavior
minimal	simple lookup	terse fragments, often one line
terse	default coding work	drop filler; fragments OK
full	explanations, design, anything risky	normal prose

Development

npm test

Pure Node.js, zero runtime dependencies. Tests use the built-in node:test runner.

License

MIT © Narek Keshishyan. See LICENSE.