# Short Answers

> Keep replies as short as the job allows without dropping what the user needs to act on. Use when the user says answers are too long, asks for concise, terse, caveman-style or ADHD-friendly replies, mentions output tokens or verbosity, or wants lean reply defaults in AGENTS.md or CLAUDE.md.

- Skill: `misssophiesterling/short-answers` (Agent Skill)
- Install (CLI): `npx skillmds@latest add misssophiesterling/short-answers`
- Raw SKILL.md: https://api.skillmd.com/api/skills/misssophiesterling/short-answers/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: MissSophieSterling (https://skillmd.com/u/misssophiesterling)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/misssophiesterling/short-answers

---


# Short answers

Output tokens are the priciest tokens: on current Claude models output costs five times as much as input, and other vendors price it in the same direction. A token is about three quarters of an English word, so a 100-word reply is roughly 130 tokens (about 30% more on newer Claude tokenizers). In agent sessions, though, output is usually the smaller part of the bill (the context re-read on every turn is the big one), so this skill trims real waste but doesn't replace the context habits in `quiet-tools` and `context-audit`.

## Rules

- Lead with the result or the answer. No preamble, no restating the request, no "I'll now…".
- For a code change: what changed and where (`file:line`), plus one line of why if it isn't obvious. Don't narrate the diff.
- No closing recap, no "let me know if…", no menu of optional extras.
- Lists only when they save words; tables only when comparing.
- Keep exact values, commands, paths and error text. Short never means vague.
- Expand when asked, when the user has to decide something, or when there's a risk they need to know about.
- Plain words over shorthand. Caveman-style grammar ("why use many token") trims a little more, but an independent test on real agent work measured only single-digit savings, because code and tool calls dominate the output. Use it only if the user asks for it.

## Reasoning is output too

Hidden reasoning is billed as output. For mechanical work, a lower reasoning effort saves more than trimming the visible reply:

- Claude Code: `/effort low` (up to `max`) for the session.
- Codex: the effort picker in `/model`, or `model_reasoning_effort` in `~/.codex/config.toml`.
- OpenAI API: `reasoning.effort`, plus `text.verbosity` (`low`, `medium`, `high`) for the length of the visible answer.

Changing effort or model partway through a session can throw away the prompt cache on some hosts (Claude Code documents this), so pick it at the start.

## Make it permanent

One line for the instruction file (AGENTS.md, CLAUDE.md or GEMINI.md):

```
Keep responses as short as the task allows: lead with the result, skip preamble, recaps and closing offers, and expand only when asked.
```

Claude Code also ships a built-in style for this: `/output-style concise`. It leads with the result, skips narration, and still keeps error reports, security warnings and confirmations for destructive actions complete.

Codex already defaults its current models to low verbosity. `model_verbosity = "low"` in `~/.codex/config.toml` sets it explicitly (it applies to Responses-API providers).

## Before you finish

- [ ] The first sentence answers the question or states the result.
- [ ] No sentence only restates, previews or offers.
- [ ] Every command, path and error the user needs is still there.

---
Part of [agent-skills](https://github.com/MissSophieSterling/agent-skills) · MIT · improvements welcome as issues.