免费在 GitHub 公开

gate-check

Find the decisions in a pipeline that do not need the expensive model and propose the gate for each: a rule, a classic classifier, or a small model, with fail-closed routing. Use when the user asks to cut model costs or latency, says the big model handles everything, or wants a triage / routing / filter layer in front of an agent. Analysis plus an optional baseline scaffold. Do NOT use for auditing context layout (context-audit) or for building eval suites (evals-bootstrap).

中文简介正在准备中。当前显示原文说明。

undefined-ui/second-brain-os

查看安装方法

包含的文件（1）

SKILL.md3.4 KB

SKILL.md（原文）

安装前可以查看交给 AI 代理的指令内容。

Find the gates

Theory: Cheap decisions and Gate practice. An agent does two kinds of work: it writes, which needs a big model, and it decides — is this spam, which queue, does this need a person — which is a bounded question the answer to which comes from a set you already know. A gate is a cheap decision layer that sorts the stream so the expensive model only sees the items that actually need judgement.

Workflow

Map the decisions. Read the pipeline's entry points and prompts and list every decision made before or during a model call: classification, routing, filtering, yes/no triage, priority, language, "is this even for us". Ignore the writing — only bounded decisions with a known label set.
Count what each costs today. For each decision currently made by the big model: calls per day if known, tokens per call, and what a wrong answer costs. A decision worth one bit that burns a frontier call is the headline finding.
Propose the cheapest gate that can hold it, in rising order of cost:
- a rule — regex, allowlist, header check: free, instant, blind to anything unanticipated; always the first layer, never the last
- a classic classifier — logistic regression or similar over simple features, trained on a few hundred labelled examples: milliseconds, fractions of a cent, and the baseline every fancier option must beat
- a small / System One model — when the input is too varied for features but the output is still a label with a confidence score
Route fail-closed. Every gate needs a confidence threshold, and doubt goes down the safe path: unsure means escalate to the big model (or a person), never means guess. Say explicitly what each gate's unsure route is.
Offer the baseline scaffold. If the user wants to proceed, generate the module's thirty-minute exercise for their data: a label.py that samples ~200 real examples for hand-labelling, and a baseline.py that trains the classic classifier and prints accuracy against a held-out split. Every vendor claim and small-model option must beat this number on their data before it earns a place in the pipeline.

Output format

Gate check — <pipeline>
decisions found: <n>, currently on the big model: <n>

1. <decision> — <where in the code>
   today: <who decides, est. cost>   label set: <the labels>
   gate: <rule | classifier | small model> — <why this tier>
   unsure -> <escalation path>
   saves: <est. calls/tokens diverted>
...

Rank by savings. If a decision genuinely needs the big model — open-ended, no stable label set, wrong answers are cheap to fix — say so and leave it alone; a gate that guesses is worse than no gate.

评价

还没有评价。欢迎分享你的使用感受。

无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

同一仓库的技能

context-audit

免费

Audit an agent's context layout against the four places: system prompt, tools, history, tail. Use when the user asks to audit, review or fix their agent's context, prompt caching, token spend per turn, or CLAUDE.md / AGENTS.md layout. Read-only: reports findings and a fix list, changes nothing. Do NOT use for writing evals (evals-bootstrap) or for loop design (goal-test).

中文简介正在准备中。当前显示原文说明。