Token Budget Advisor

核心功能：拦截回答流程，提供4档（25%-100%）回答深度选项，支持快捷指令，会话内保留用户选择，精准控制token用量与细节。典型场景：用户提及token、预算、深度或short/tldr等关键词，想提前控细节时触发（已选深度或答案极短则不触发）。输入为用户需求，输出先展示含token估算的深度菜单，再按选择输出对应内容。

djc00p

热度 45.24

最近更新 2026-05-11 17:40:23

分类：效率依赖：纯提示词应用场景：摘要总结 / 模板

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add token-budget-advisor

Metadata

name: token-budget-advisor
description: Offer users informed choice about response depth before answering. Control token usage, response length, and answer depth. Trigger phrases: token budget, token count, token usage, response length, short version, brief answer, detailed answer, exhaustive, save tokens. Adapted from everything-claude-code by @affaan-m (MIT)
metadata: {"clawdbot":{"emoji":"💰","requires":{"bins":[],"env":[]},"os":["linux","darwin","win32"]}}

Token Budget Advisor

Intercept response flow to let users choose depth before answering.

When to Use

User mentions tokens, budget, depth, or response length
User says "short version", "tldr", "brief", "exhaustive", etc.
User wants to control detail level upfront

Do not trigger when: user already set depth this session, or answer is trivially one line.

Depth Options

Present before answering:

Input: ~[N] tokens | Type: [type] | Complexity: [level]

[1] Essential (25%)   -> ~[tokens]   Direct answer only
[2] Moderate (50%)    -> ~[tokens]   Answer + context + example
[3] Detailed (75%)    -> ~[tokens]   Multiple examples + alternatives
[4] Exhaustive (100%) -> ~[tokens]   Everything

Level Guide

Level	Length	Include	Omit
25%	2-4 sent	Direct answer, conclusion	Context, examples
50%	1-3 para	Answer + context + 1 example	Deep analysis, edge cases
75%	Structured	Examples, pros/cons, alternatives	Extreme cases
100%	No limit	Everything	Nothing

Shortcuts

If user signals depth, respond at that level immediately:

"1" / "25%" / "short" / "brief" / "tldr" → 25%
"2" / "50%" / "moderate" → 50%
"3" / "75%" / "detailed" → 75%
"4" / "100%" / "exhaustive" → 100%

Maintain user's chosen level for rest of session unless changed.

Estimation

Heuristic-based (no real tokenizer):

Input tokens = estimate from prompt size
Response window = input × complexity multiplier
Accuracy ~85-90%, variance ±15%