Give every idea a nightly AI co-founder

The problem

Ideas do not die because they are bad. They die because evaluating one properly takes hours: competitor research, demand signals, pricing, an honest look at whether you are the right person to build it. Nobody does that for a note, so the note app fills up and you keep chasing whichever idea felt exciting last night.

I had 36 ideas in my vault and another 210 buried in an old Notion brain dump. Zero had ever been researched.

What I built

Idea Forge: a private repo where every idea is one markdown file, plus a scheduled Claude Code job (launchd on my Mac mini, 9am daily) that acts as a co-founder. Every run it:

1. Web-researches each new or edited idea, with cited sources.
2. Scores it 0 to 100 on a 9-dimension rubric matched to my founder profile.
3. Tags it Build now, Explore, Park, or Pass.
4. Writes a mini PRD for anything scoring 65 or higher.
5. Rebuilds the leaderboard and leaves me a short co-founder note.

Capture costs one line. Anywhere in my Obsidian vault or in Telegram I type:

idea: meal prep planner for shift workers

A dumb script (no LLM) picks it up, dedupes it, and creates a stub file. The agent does the rest on the next run. Ideas get re-researched when I edit the pitch or after 21 days, and the current top 5 get a deeper pass.

First pass: 34 of 36 ideas rated. The top pick scored 77 and already had a PRD waiting. 19 ideas I was quietly attached to got an honest Pass, with sources.

Copy the pattern

You need one folder, two files, and a schedule.

1. An ideas/ folder, one markdown file per idea, however rough.
2. A rubric.md (the dimensions you score on: market size, distribution, your unfair advantage) and a short founder-profile.md, so the agent scores for you, not for a generic founder.
3. A brief the agent follows every run. Mine boils down to:

You are my product co-founder. For each idea file with no score,
or edited since the last run:
1. Research the market with web search. Cite every claim.
2. Score it 0 to 100 against rubric.md and founder-profile.md.
3. Categorize it: Build now, Explore, Park, or Pass.
4. If it scores 65 or higher, write a one-page PRD next to it.
Then rebuild LEADERBOARD.md and write CO-FOUNDER-NOTE.md:
what changed, what I should build next, and why.

4. Schedule it. I use launchd with a few guards (lock file, quiet hours), but plain cron gives you the same loop:

0 9 * * * cd ~/idea-forge && claude -p "Run the co-founder brief in CLAUDE.md"

The point is not the tooling. It is the split: capture stays free, diligence gets scheduled, and you only ever look at a ranked list. It is the same scheduled-agent move behind putting your competitor research on autopilot: let the slow work run overnight so you wake up to answers, not a to-do list.

Do this now

1. Make an ideas/ folder and dump every idea into it, one file each.
2. Save the brief above as CLAUDE.md in that folder, next to your rubric and founder profile.
3. Schedule the daily run, then check the leaderboard tomorrow.