Autosearch autosearch:router

Route an autosearch task to the smallest relevant group of skills, then load only the matching leaf skills. First line of defense against loading all 80+ skill SKILL.md files at session start.

install

source · Clone the upstream repo

git clone https://github.com/0xmariowu/Autosearch

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/0xmariowu/Autosearch "$T" && mkdir -p ~/.claude/skills && cp -r "$T/autosearch/skills/router" ~/.claude/skills/0xmariowu-autosearch-autosearch-router && rm -rf "$T"

manifest: autosearch/skills/router/SKILL.md

source content

Routing Policy

When the runtime AI picks up an autosearch research task, do not try to read every leaf

SKILL.md

. Follow this three-step routing:

Identify 1–3 groups from the query intent. Use the keyword hints below as a first pass; use the domain + scenario tags on each group index for ambiguity.
Read only those group indexes (
```
references/groups/<group>.md
```
). Each group index lists its leaf skills with one-line triggers and suggested model tier.
Pick 3–8 leaf skills from the matched groups. Read their
```
SKILL.md
```
only when you're about to call them.

Do not enumerate every leaf skill to the main model. That defeats the entire progressive-disclosure design and burns tokens.

Group Selection Hints

Content surface

If the query mentions…	Pick group
小红书 / 抖音 / B站 / 微博 / 知乎 / 播客 / 快手 / 雪球 / V2EX	`channels-chinese-ugc`
36kr / CSDN / 掘金 / InfoQ 中文 / 微信公众号	`channels-cn-tech`
paper / arxiv / citation / benchmark / 论文 / survey / openreview	`channels-academic`
github / repo / code / issue / npm / pypi / huggingface	`channels-code-package`
crunchbase / 融资 / producthunt / G2 review	`channels-market-product`
stack overflow / hacker news / dev.to / reddit	`channels-community-en`
twitter / X / linkedin / 职业 / career	`channels-social-career`
一般网页 / 搜索引擎 / tavily / exa / ddgs / searxng / rss	`channels-generic-web`
视频 / 字幕 / 转录 / podcast / youtube	`channels-video-audio`

Tool surface

If the task is…	Pick group
Fetch a URL, render JS, run interactive browser, download media	`tools-fetch-render`
Clarify user intent, decompose a task, recall known info, gene queries	`workflow-planning`
Normalize, rerank, anti-cheat, extract dates, score with LLM, apply rubrics	`workflow-quality`
Assemble context, extract knowledge, build knowledge map, synthesize report	`workflow-synthesis`
Track outcomes, auto-evolve skills, create new skills, capture/compact experience	`workflow-growth`

Model Tier Escalation

Every group index and leaf skill carries a

model_tier

field (

Fast

Standard

Best

). Default router execution is

Fast

. Escalate only when the group index or leaf skill calls for a higher tier on the specific step (for example,

synthesize-knowledge

needs

Best

;

search-bilibili

Fast

Rule of thumb:

Fast: retrieval, normalization, schema checks, URL reading — the bulk of every session.
Standard: semantic ranking, evidence extraction, source curation, mid-complexity planning.
Best: clarify, decompose, synthesize, evaluate delivery, skill evolution — the 1–2 steps that shape the whole outcome.

See

autosearch:model-routing

(if available in the runtime) for the full advisory catalog.

What the Router Does Not Do

Does not call any leaf skill itself — it only points.
Does not guess: if no group matches, ask the user or degrade to
```
channels-generic-web
```
.
Does not cache across sessions — skill state lives in each skill's
```
experience.md
```
(read lazily).

Quality Bar

Evidence items have non-empty title and url.
No crash on empty or malformed API response.
Source channel field matches the channel name.