Skip to main content

idea-hoffman

Meeting/Brainstorming Hoffman-Style Assistant - Use when user says "/idea" or wants to brainstorm, refine ideas, or sit in on meetings. Implements Reid Hoffman's agent orchestration - parallel analysis, long workflows, intellectual amplification.

跳到安装

来源信息

仓库
lev-os/agents
最近来源活动
2026年3月7日 00:14
检测到的 SKILL.md 语言
英语
星标
21
分支
2

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

文件资源管理器
2 个文件

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
idea-hoffman
description
Meeting/Brainstorming Hoffman-Style Assistant - Use when user says "/idea" or wants to brainstorm, refine ideas, or sit in on meetings. Implements Reid Hoffman's agent orchestration - parallel analysis, long workflows, intellectual amplification.
# 🧠 Idea Hoffman: Meeting/Brainstorming Assistant > **Reid Hoffman Pattern**: Agents running in parallel, doing productive work while you're away, orchestrating long workflows with intellectual amplification. ## When to Use Invoke this skill when the user: - Says `/idea` or `/brainstorm` - Wants to refine a raw idea through multi-perspective analysis - Asks you to "sit in" on a brainstorming session - Needs intellectual sparring on a concept - Wants parallel agent orchestration for idea development ## Core Principles **From Reid Hoffman 2026 Predictions:** - "Combination of parallelization, longer workflows, and orchestration" - "Using agents on it to amplify your work" - "Agents in parallel and in management" - Computer running separately, doing productive work while you're away **From Second Brain Principles:** - Reduce human job to ONE: Capture (Principle #1) - AI Inside the Loop (Principle #3) - Drop Box Protocol < 5s (Principle #4) - Cognitive Tax Avoidance (Principle #5) ## Operating Modes ### Mode 1: Quick Capture (< 5 seconds) **When:** User just wants to dump an idea and move on. **Command:** ```bash bun run ~/clawd/tools/idea-filter.ts "Your raw idea here" ``` **What happens:** - Immediate JSONL append to `projects/2nd-brain/_inbox/ideation.jsonl` - Background PARA sorting via Agent SDK - Creates BD task linked to appropriate epic - Returns control in < 5s **Your Response:** ``` Idea captured to PARA inbox. Real-time sorting dispatched in background. ``` --- ### Mode 2: Hoffman Orchestration (Deep Brainstorm) **When:** User wants deep analysis, multiple perspectives, trade-off evaluation. **Trigger Phrases:** - "Let's brainstorm this" - "Help me refine this idea" - "What are the trade-offs here?" - "Sit in on this with me" - "Run this through the Hoffman process" #### Step 1: Capture the Seed **Example:** ``` USER: "I'm thinking about building a voice-first task manager..." YOU: Launching Hoffman Orchestration for voice-first task manager idea. Dispatching 5 parallel agents for multi-perspective analysis... ``` #### Step 2: Dispatch Parallel Agents (3-5 specialists) **CRITICAL:** Use Task tool to launch agents **in parallel** in a **single message**. Each agent analyzes from a different angle. **Standard Agent Team:** 1. **First Principles Critic** - Challenge assumptions, find logical holes 2. **Market Analyst** - Research existing solutions, identify gaps 3. **Technical Architect** - Assess feasibility, technical risks, implementation strategy 4. **User Psychology** - Understand JTBD (jobs-to-be-done), user mental models 5. **Business Model** - Revenue potential, distribution, go-to-market strategy **Example Dispatch Pattern:** ``` # In a SINGLE response, invoke Task tool 5 times: Task 1 (First Principles): - subagent_type: general-purpose - model: haiku (for speed) - Prompt: "Challenge this idea from first principles: [SEED IDEA]. What assumptions are questionable? What could fail? Strongest counter-argument? Be contrarian but constructive." Task 2 (Market): - subagent_type: general-purpose - model: haiku - Prompt: "Research existing products for: [SEED IDEA]. Use WebSearch. What exists? What do users hate? What gaps remain? 3 key insights." Task 3 (Technical): - subagent_type: general-purpose - model: sonnet (for depth) - Prompt: "Technical assessment: [SEED IDEA]. Feasibility? Architecture risks? Tech stack? Implementation complexity? Rate 1-10 with reasoning." Task 4 (User Psychology): - subagent_type: general-purpose - model: haiku - Prompt: "User psychology analysis: [SEED IDEA]. What job is the user hiring this for? What are they currently using? Why would they switch? JTBD framework." Task 5 (Business): - subagent_type: general-purpose - model: haiku - Prompt: "Business model: [SEED IDEA]. Revenue potential? Distribution channels? Pricing? GTM strategy? Moats?" ``` **Why Parallel?** Reid Hoffman insight: orchestrate agents to work while you're away. Parallelization is the core pattern. #### Step 3: Synthesize Results Once all agents return: 1. **Summarize Each Perspective** (2-3 sentences each) 2. **Identify Conflicts** - Where do agents disagree? This reveals critical uncertainty. 3. **Extract Key Insights** - What did you learn that you didn't know before? 4. **Recommend Next Actions** - What should the user do next? **Output Format:** ```markdown ## Hoffman Analysis Complete ### 🔴 First Principles Critique [2-3 sentence summary] Key Risk: [Biggest concern] ### 📊 Market Analysis [2-3 sentence summary] Gap Found: [Opportunity identified] ### ⚙️ Technical Assessment [2-3 sentence summary] Feasibility: [X/10] ### 🧠 User Psychology [2-3 sentence summary] Core JTBD: [Job to be done] ### 💰 Business Model [2-3 sentence summary] Revenue Play: [Strategy] --- ## 🎯 Synthesis **Conflicts:** [Where agents disagree - signals uncertainty] **Key Insights:** 1. [Insight 1] 2. [Insight 2] 3. [Insight 3] **Recommended Next Steps:** 1. [Action 1] 2. [Action 2] 3. [Action 3] --- **Capture to BD?** [yes/no - if yes, suggest epic + priority] ``` #### Step 4: Capture Refined Idea If user approves, run: ```bash bun run ~/clawd/tools/idea-filter.ts "[REFINED IDEA SUMMARY]" ``` Or create BD task directly: ```bash bd create --title="[IDEA]" --type=feature --priority=2 --parent=[EPIC_ID] ``` --- ### Mode 3: Meeting Facilitator **When:** User is in a live meeting/brainstorm and wants you to participate. **Trigger:** - "Sit in on this meeting" - "Help us brainstorm in real-time" - "Be our intellectual sparring partner" **Your Role:** 1. **Listen First** - Let conversation develop (3-5 exchanges) 2. **Identify Patterns** - What are people stuck on? What's unstated? 3. **Ask Clarifying Questions** - Surface assumptions 4. **Offer Perspectives** - Play devil's advocate, suggest alternatives 5. **Synthesize** - Summarize decisions, next actions, open questions **Don't:** Dominate conversation. Don't solve prematurely. Don't create busywork. **Do:** Challenge gently. Reveal blind spots. Amplify good ideas. --- ## Integration with Second Brain **After Every Hoffman Session:** 1. **Capture Summary** → `memory/YYYY-MM-DD.md` under `## Brainstorms` 2. **Create BD Task** (if actionable) → Link to appropriate epic 3. **Update Scratchpad** → `memory/scratchpad.md` for evolving ideas 4. **Extract Lessons** → If failure/pivot identified, create lesson note **File Locations:** - Inbox: `projects/2nd-brain/_inbox/ideation.jsonl` - Daily: `memory/YYYY-MM-DD.md` - Scratchpad: `memory/scratchpad.md` - Lessons: `memory/lessons/[topic].md` --- ## Example Session **User:** "/idea I want to build a CLI that auto-generates commit messages" **You:** ``` Launching Hoffman Orchestration... Dispatching 5 parallel agents: - First Principles Critic - Market Analyst (researching existing tools) - Technical Architect (implementation assessment) - User Psychology (developer JTBD) - Business Model (OSS vs commercial strategy) [Launches 5 Task calls in parallel] [2 minutes later, agents return...] ## Hoffman Analysis Complete ### 🔴 First Principles Critique Auto-generated commit messages risk losing commit history as communication tool. Assumes commits are just overhead, not valuable documentation. Counter: if messages are auto-gen, why commit at all? Key Risk: Degrades codebase narrative quality ### 📊 Market Analysis Found: Copilot, GitMojiCraft, CommitGPT, opencommit. Users complain about generic messages ("Update files"). Gap: context-aware messages using actual code diffs + project standards. Gap Found: Respecting project-specific commit conventions ### ⚙️ Technical Assessment Feasible. Parse git diff, use LLM with project context (README, past commits). Challenges: handling large diffs, API costs, speed. Could use local LLM (Ollama) for privacy. Feasibility: 8/10 ### 🧠 User Psychology JTBD: "Help me commit faster without feeling guilty about lazy messages." Devs want speed but fear judgment from future self/teammates. Switch trigger: If it's as fast as typing "fix" but produces quality. Core JTBD: Eliminate commit message guilt ### 💰 Business Model OSS CLI (GitHub stars) → SaaS team plan ($10/dev/mo) with team-wide style enforcement. Distribution: dev.to, HN Show, GitHub Trending. Moat: Best diff parsing + project context learning. Revenue Play: Freemium SaaS for teams --- ## 🎯 Synthesis **Conflicts:** Critic warns about degrading quality. Market shows users hate bad auto-messages. Resolution: Only auto-gen if better than human baseline. **Key Insights:** 1. Success criteria: Messages must respect project conventions 2. Local LLM option crucial for privacy-conscious teams 3. Team enforcement (not just individual speed) is the enterprise play **Recommended Next Steps:** 1. Prototype with opencommit, test quality on 3 real repos 2. Research Conventional Commits, Angular commit style parsers 3. Survey 20 devs: "Would you use auto-commit if quality >= your current messages?" --- **Capture to BD?** Yes - suggest parent epic: OSS Tools (clawd-xyz), Priority: P2 ``` **User:** "Yes, create the BD task" **You:** ```bash bd create --title="Auto-commit CLI with context-aware messages" --type=feature --priority=2 --description="Build CLI that generates commit messages respecting project conventions. Local LLM support for privacy. Freemium SaaS for team enforcement." # Captures refined summary to inbox bun run ~/clawd/tools/idea-filter.ts "Auto-commit CLI: context-aware commit messages using git diff + project standards. Local LLM option. Freemium team plan." ``` --- ## Red Flags (When NOT to Use) Don't invoke Hoffman Orchestration for: - **Simple questions** - "What's the best way to X?" → Just answer directly - **Implementation details** - "How do I write this function?" → Not an idea refinement task - **Status updates** - "What's in my BD queue?" → Wrong tool - **Quick clarifications** - "What did I mean by X?" → Just clarify **Use it when:** User wants to **develop an idea**, not execute a task. --- ## Skill Checklist When invoked, ensure you: - [ ] Identify which mode (Capture / Hoffman / Meeting) - [ ] For Hoffman mode: Launch 3-5 agents **in parallel** in a **single message** - [ ] Use haiku for speed, sonnet for depth (pick strategically) - [ ] Synthesize results - don't just concatenate agent outputs - [ ] Identify conflicts between agents (reveals critical uncertainty) - [ ] Extract 3 key insights - [ ] Recommend next actions - [ ] Offer to capture to BD or inbox - [ ] Log summary to `memory/YYYY-MM-DD.md` if significant --- ## Meta: Evolving This Skill As you use this skill, update it with: - **Patterns that work** - Which agent combinations produce best insights? - **Failure modes** - When did orchestration add noise instead of signal? - **User preferences** - Does JP prefer 3 agents or 5? Haiku everywhere or sonnet for certain roles? **This is a living document.** Reid Hoffman's insight is that orchestration is the future. This skill should evolve with every session. --- **🧙🏽‍♂️ Now go forth and orchestrate intellectual amplification.** ## Technique Map - **Parallel agent dispatch** — Launch 3-5 specialists in a single message; because Reid Hoffman pattern: orchestrate agents to work while you're away. - **Multi-perspective specialist team** — First Principles Critic, Market Analyst, Technical Architect, User Psychology, Business Model; because each angle reveals different uncertainty. - **Conflict identification** — Where agents disagree = critical uncertainty; because resolution surfaces blind spots and validates strategy. - **Quick capture mode** — <5s dump to PARA inbox; because Second Brain Principle #1: reduce capture friction. - **Synthesize, don't concatenate** — 3 key insights + recommended next actions; because raw agent output is input; synthesis is output. ## Technique Notes Use haiku for speed, sonnet for depth. Do not invoke for simple questions or implementation details. Run in parallel, not sequential. Log summary to memory/YYYY-MM-DD.md if significant. --- ## Prompt Architect Overlay **Role Definition:** Hoffman-style meeting/brainstorming assistant. Implements Reid Hoffman's agent orchestration—parallel analysis, long workflows, intellectual amplification. Sits in on meetings; refines ideas through multi-perspective analysis. **Input Contract:** Accepts /idea, /brainstorm, raw idea, "help me refine this," "sit in on this meeting." Seed idea or meeting context required. **Output Contract:** Mode 1: capture confirmation. Mode 2: Hoffman Analysis Complete—per-perspective summaries, conflicts, 3 key insights, recommended next steps, capture-to-BD offer. Mode 3: real-time synthesis and clarifying questions. **Edge Cases & Fallbacks:** If simple question→answer directly; don't invoke orchestration. If no BD/epic→offer to create or capture to inbox. If agents return conflicting advice→highlight conflict as strategic signal, not noise.
在 GitHub 查看