Why waste time say lot word when few word do trick?
I kept watching Claude narrate every action like a sports commentator. "I'll read the file now." "Let me check the configuration." "Here's a summary of what I did." That's tokens I'm paying for and attention I'm losing.
So I turned Kevin Malone's philosophy into a Claude Code skill that cuts ~35-45% of output tokens — dropping filler prose and low-value narration while keeping the explanations that actually matter.
npx skills add ajmalhassan/kevin-modeOther install methods
Clone into skills directory:
git clone https://github.com/ajmalhassan/kevin-mode.git ~/.claude/skills/kevin-modeOr copy just the skill file:
mkdir -p ~/.claude/skills/kevin-mode
cp SKILL.md ~/.claude/skills/kevin-mode/| What gets cut | What stays |
|---|---|
| Articles, filler pronouns, transition words | Non-obvious why explanations |
| "Let me X" announcements | Genuine questions |
| Trailing summaries | Risk/warning flags |
| Restating what you just said | All code output |
| Status updates for obvious actions | Commits, PRs, docs (others read those) |
Explaining a change:
Normal: "I'll read the configuration file to understand how the routes are set up. I can see the middleware chain uses express.Router(). I'll add the new endpoint after the existing auth routes. The POST handler needs to come before the catch-all route so it doesn't get swallowed."
Kevin: "POST handler before catch-all — catch-all swallows it otherwise."
Asking a question:
Normal: "We could add validation in the middleware layer or directly in the handler function. Middleware catches invalid requests earlier. Which approach would you prefer?"
Kevin: "Validation — middleware or handler? Middleware catches earlier."
Completing a task:
Normal: "I've finished implementing the changes. I updated the parser, added validation, and updated the tests. All 12 pass. Let me know if you'd like adjustments."
Kevin: "Done. Parser, validation, tests. 12 pass."
Invoke directly:
/kevin-mode
| Output type | Reduction |
|---|---|
| Prose tokens | ~60-70% |
| Code tokens | ~0% (unchanged) |
| Tool calls | ~5-10% (fewer confirmation rounds) |
| Net output | ~35-45% |
You could paste "be terse" into your CLAUDE.md. But it drifts — Claude gradually returns to its verbose defaults mid-session. A skill is loaded fresh on invocation and provides concrete examples that calibrate the style precisely.
MIT