Research Spinoff

Research Spinoff — Multi-Agent Evaluation Chain

Take an existing project/tool/concept and discover what other products could be spun out from it.

Workflow

Step 1: Parallel Research (2 agents, GLM-5)
   ├── Agent A: Surfaces, audiences, domains, monetization
   └── Agent B: AI-native, creative, non-obvious, dev tools
         │
Step 2: Evaluate + Rank (1 agent, GLM-5)
         │ dedupe, score on uniqueness × speed × revenue
         │ output: top 10 ranked
         │
Step 3: Competition Scout (1 agent, GLM-5)
         │ for each top 10: search GitHub, ProductHunt, AppStore
         │ tag: "blue ocean" / "crowded" / "few competitors"
         │
Step 4: Lattice Frame (local, no LLM)
         │ lattice suggest on top 5
         │ attach relevant mental models to each decision
         │
Step 5: Post results to thread
Step 6: User picks → auto-scaffold project

Execution

Step 1: Spawn Research Agents (parallel)

Spawn 2 sub-agents with runtime: subagent, model: GLM, mode: run.

Agent A prompt — see references/prompt-agent-a.md Agent B prompt — see references/prompt-agent-b.md

Replace {{PROJECT}} and {{CONTEXT}} placeholders with the actual project name and description.

Step 2: Evaluate + Rank

After BOTH research agents return, spawn 1 evaluator agent:

• Input: combined raw ideas from Agent A + Agent B
• Task: deduplicate, score each idea on a 3-axis matrix (uniqueness 1-5, build speed 1-5, revenue potential 1-5), multiply scores, rank top 10
• Output: markdown table with columns: Rank, Name, Pitch, Audience, Build Time, Revenue Model, Score
• See references/prompt-evaluator.md

Step 3: Competition Scout

Spawn 1 competition scout agent on the top 10:

• For each idea: search GitHub, ProductHunt, Google for direct competitors
• Tag each: 🟢 blue ocean (0 competitors), 🟡 few competitors (1-3), 🔴 crowded (4+)
• Add competitor names/URLs where found
• See references/prompt-scout.md

Step 4: Lattice Frame

Run locally (no sub-agent needed):

# For each top 5 idea
lattice suggest "<idea pitch>" --count 3 --no-llm --json

Attach the suggested mental models to each idea in the final output.

Step 5: Post Results

Create a Discord thread (or post to existing thread):

• Thread name: 🔬 Spin-off Product Ideas from {{PROJECT}}
• Post: ranked table + competition tags + lattice framing
• Ask user which to build

Step 6: Auto-Scaffold (on user pick)

When user picks an idea:

mkdir -p ~/github/{{project-name}}
# Generate: README.md, PRD.md from research pitch
# Init: go mod init / npm init / etc based on stack
# Copy: CI template from references/

Thread Management

• Always create a new thread in the user's current channel for results
• Keep the source thread clean — only post the final summary there
• Reference the results thread from the source

Notes

• All research agents use GLM-5 (fast, cheap, good at web search)
• Steps 1 is parallel; steps 2-4 are sequential (each depends on previous)
• Total wall-clock time: ~8-12 minutes
• If lattice is not installed, skip Step 4 gracefully