Idea review agent
An agent whose job is to talk me out of bad ideas
An orchestrator agent captures ideas, researches them, and runs a product-manager review that ends in a Pursue, Park, or Pass verdict for a person to decide.
How it fits together
Inputs
An idea by message
A sentence sent from a phone to the orchestrator's topic.
Telegram
Engine
Orchestrator
Keeps the idea queue, projects, and tasks in Postgres, and prepares work for a person to start.
- Claude
Supabase
Parse by rules first
A deterministic parser reads each message before any model, so a command is never misread as an idea.
Clarify
A new idea gets a few clarifying questions before any research starts.
- Claude
Research
Bounded web research stores findings and a ready kickoff prompt with the idea.
- Claude
Product-manager review
Demand, fit, feasibility, economics, timing, and the assumptions that would kill it, scored on a weighted scorecard with a confidence rating.
- Claude
Outputs
Viability brief
Pursue, Park, or Pass, the cheapest experiment that would test it, and what evidence would change the call.
A person decides
The brief is a draft for review. The agent recommends and never commits to a project.
Guardrails
- Every market claim cites a source or is labeled inference
- Every brief names the strongest argument against the idea
- The orchestrator prepares work and never launches builds or sends messages
The problem
A founder has more ideas than time, and the expensive mistake is building the wrong one. A model that only helps an idea along makes that mistake easier. The review needed to be skeptical by design and cheap enough to run on every idea.
What I built
- Idea capture by text message, with clarifying questions and server-side research before anything is scored.
- A product-manager skill whose stated job is to kill bad ideas cheaply, with a weighted scorecard and fixed bands for Pursue, Park, and Pass.
- An honesty rule: market claims carry a source or an inference label, and a low-confidence high score lands in Park, not Pursue.
- A queue the review works through, writing each brief back as a draft for a person.
- A monthly review that gives every live automation a keep, convert, or retire verdict.
How I built it
Intake runs in n8n behind the shared Telegram router, with Claude for clarifying questions and research and Postgres on Supabase for the queue. The review itself is a Claude Code skill, so the same written method runs every time and can be read and changed like code.
What it shows
I build agents that argue with me when they should, and I keep the decision with a person.