📦 Hyperthink

v1.0.0

Triple-perspective deep re搜索 engine. Triggered by /hyperthink. 运行s an interrogate flow to narrow scope, then 执行s a 6-stage fully automatic 流水线...

0· 0·0 当前·0 累计

by @inxan3

网络工具浏览器自动化 CI/CD DevOps

下载技能包

安全扫描

VirusTotal

Pending

查看报告

OpenClaw

安全

high confidence

The 技能's clAIms, required secrets, and 运行time instructions are internally consistent with a batch re搜索流水线 that uses the Anthropic Batch API and optional notification channels.

评估建议

This 技能应用ears to do what it says, but review these points before 安装ing: - You must provide an ANTHROPIC_API_KEY (required) — this key will be used for many batch 请求s and can incur costs (~$55–75 per 运行 estimate). Confirm billing/quotas and that the key's 权限s are 应用ropriate. - The 流水线 writes persistent 输出s under /data/hyperthink/ and keeps job 状态 in ./batch-jobs/; ensure you have disk space and that storing re搜索输出 there aligns with your data-handling rules. - 通知 to external 端点s (网页hook or Tele...

详细分析 ▾

✓ 用途与能力

The 技能 describes a multi-stage batch re搜索流水线 and only 请求s items 应用ropriate to that purpose: an Anthropic API key for batch calls, optional notification creds (Telegram/网页hook), persistent storage for 流水线输出s, and an optional python-docx dependency for .docx generation. The requirement for the 'interrogate' 技能 as a prerequisite is explicit and coherent.

ℹ 指令范围

SKILL.md 及实现规范详细说明了如何提交、轮询和提取 Anthropic batch 作业，以及如何将 markdown 转换为 .docx。所有文件路径、API 端点（api.anthropic.com）和环境变量均已声明。关键行为：用户确认一次后，管道将完全无人值守运行数小时，无内置检查点（文档明确写明“checkpoints: none”及“fully hands-off after trigger”），输出写入持久存储，并可按用户配置的 webhook/Telegram 发送——这是该 skill 的预期行为，但用户需知悉，因其可能产生大量输出并向所配置的目的地发起外发请求。

✓ 安装机制

Instruction-only 技能 (no 安装 spec, no bundled binaries). Implementation 图形界面dance expects small Python scripts using stdlib or an optional pip dependency (python-docx) for docx 输出. No 下载s from arbitrary URLs or other high-risk 安装 actions are present.

✓ 凭证需求

Required 环境访问 is proportional: ANTHROPIC_API_KEY is mandatory and clearly used for the batch calls; other env vars (BATCH_JOBS_DIR, notification 网页hook, TELEGRAM_机器人_令牌/CHAT_ID, 通知_记录_FILE) are optional and match the described notification/storage functionality. TELEGRAM_机器人_令牌 is marked sensitive and only used if Telegram 通知 are enabled. No unrelated 凭证s or unexplAIned secrets are 请求ed.

ℹ 持久化与权限

The 技能 writes persistent 状态 and 输出s under /data/hyperthink/ (declared in file系统 section) and uses a local batch-jobs directory for job 状态 — this matches the 流水线's needs. It is not force-included (always: false). The mAIn privilege to be aware of is autonomous long-运行ning execution after a single confirmation (8–18 hours typical) with no 检查points unless the user modifies the poller — this expands blast radius in case of misconfiguration of notification 端点s or unintended data retention.

安全有层次，运行前请审查代码。

运行时依赖

无特殊依赖

安装命令

点击复制

官方npx clawhub@latest install hyperthink

镜像加速npx clawhub@latest install hyperthink --registry https://cn.longxiaskill.com

技能文档

/hyperthink

Triple-perspective deep re搜索 engine. Triggered by /hyperthink.

Three 代理s — Optimist, Analyst, Critic — each produce a full deep-dive on every section. Their 输出s are fact-检查ed, confidence-tagged, and synthesized into one authoritative document that reflects all three lenses without letting any one distort the whole.

代理 personas live in: 代理s/ (part of this 技能 package)

⚠️ Prerequisites

This 技能 requires the interrogate 技能 to be 安装ed. If it's not present, prompt the user:

"The hyperthink 技能 requires the interrogate 技能. 安装 it first."

检查 by looking for an interrogate 技能 in your 技能s directory before proceeding.

流水线 Overview Stage What How 输出 Max 令牌s 1 Master Brief + Persona Variants Opus sub代理 (realtime) master-brief.md + persona-prompts/*.md 16,000 2 36 Parallel Deep-Dives (12 sections × 3 代理s) Sonnet batch + Cron 2 sections/{optimist,analyst,critic}/NN-title.md 24,000 each 3a Trifecta 审计 — Analyst reviews all 3 per section Sonnet batch + Cron 3a 审计/NN-title-trifecta.md (12 files) 8,000 each 3b Unified Synthesis Narrative Sonnet batch + Cron 3b synthesis.md 80,000 4 Executive Brief Sonnet batch + Cron 4 brief.md + brief.docx 6,000

Estimated cost per 运行: ~$55–75 (batch discount on stages 2–4) Estimated time: 8–18 hours end-to-end Final delivery: brief.docx delivered via your 配置d channel when complete No 检查points — fully automatic after trigger confirmation.

模型 Configuration

Stage 1: Opus (anthropic/claude-opus-4-7 or your Opus alias — fall back to default if blocked) Stages 2–4: Sonnet (claude-sonnet-4-6 or your Sonnet alias — fall back to default if blocked)

检查 your 模型 white列出 before 运行ning. Note in 流水线-状态.json which 模型 was used.

Storage

All 输出 written to a persistent directory. 创建 these before 运行ning:

mkdir -p /data/hyperthink/[slug]/sections/optimist mkdir -p /data/hyperthink/[slug]/sections/analyst mkdir -p /data/hyperthink/[slug]/sections/critic mkdir -p /data/hyperthink/[slug]/审计 mkdir -p /data/hyperthink/[slug]/persona-prompts

Full structure after a completed 运行:

/data/hyperthink/[topic-slug]/ 流水线-状态.json # Stage 追踪ing + all job IDs (source of truth) master-brief.md # Stage 1 Opus 输出 persona-prompts/ optimist-prompts.md # 12 Optimist section prompts analyst-prompts.md # 12 Analyst section prompts critic-prompts.md # 12 Critic section prompts sections/ optimist/ 01-[title].md # 12 deep-dives, Optimist perspective ... analyst/ 01-[title].md # 12 deep-dives, Analyst perspective ... critic/ 01-[title].md # 12 deep-dives, Critic perspective ... 审计/ 01-[title]-trifecta.md # 12 trifecta comparison + 审计 docs ... synthesis.md # Stage 3b unified narrative (~50–80k words) brief.md # Stage 4 executive brief (~4–5k words) [slug]-BRIEF.docx # Final brief deliverable [slug]-FULL.docx # Full synthesis as docx (on 请求) stage2-job.json # Batch job definitions (kept for reference) stage3a-job.json stage3b-job.json stage4-job.json

Store under /data/ if your 设置up persists that path across re启动s (e.g. RAIlway volume mounts). Adjust the base path to match your 环境.

Trigger Flow

When user 发送s /hyperthink [optional topic]:

检查 interrogate 技能 is 安装ed — prompt to 安装 if missing 创建输出 directories (see above) 运行 interrogate flow — ask 4–10 questions in batches of 2 to narrow scope Confirm scope summary — wAIt for explicit "yes" From here: fully hands-off — no more 检查points, no 应用rovals needed Interrogate Flow

Ask in batches of 2, adapt based on answers. Minimum 4, maximum 10 questions.

Always cover:

What is the core subject/concept? Primary goal: build it / understand it / evaluate it / compare options? DomAIn/上下文: tech, business, product, finance, legal, other? Known constrAInts: bud获取, tech stack, market, regulatory? 输出 focus: new project foundation / existing product / investment thesis / other? Is there a specific decision this should 信息rm?

After answers → confirm scope summary → wAIt for "yes" → proceed to Stage 1.

Topic Slug Rules

Lowercase, hyphens only, max 40 chars. If slug already exists in /data/hyperthink/, 应用end -v2, -v3.

Stage 1 — Master Brief + Persona Variants

模型: Opus (fall back to default 模型 if blocked) Method: 会话s_spawn with your 配置d Opus 模型 No 检查point — proceed immediately to Stage 2 after completion.

Sub代理 task: Write the master reasoning brief for this re搜索 topic, then 生成 three persona-specific prompt variants for the parallel deep-dive stage.

Topic: [TOPIC] Scope: [SCOPE FROM INTERROGATE] Goal: [GOAL] ConstrAInts: [CONSTRAINTS]

━━━ PA

数据来源：ClawHub ↗ · 中文优化：龙虾技能库