AI Greenhouse Agent Tends Ideas from Seed to Ripe Content
Build a file-based AI agent that tracks ideas through 6 growth states, cross-references connections, flags ripeness via 3/5 criteria, and composts wilting ones after 14 days inactivity or 10 days without links.
Idea Growth Requires Greenhouse Conditions, Not Static Storage
Notes apps store seeds in jars, leaving them unchanged and lifeless. A greenhouse creates growth conditions: ideas evolve through six states—seed (isolated thought), signal (attached evidence), seedling (planted raw), growing (attracting connections), ripening (near writable), wilting (needing decision), and composting (retired but retrievable). This shifts ideation from reactive capture to patient tending, where files physically move between folders like seeds/, ready/, and compost/. Unlike Karpathy's LLM knowledge bases for archiving consumed info, the greenhouse grows your creative output by enforcing patience—ideas ripen after 14+ days, collecting diverse signals to reveal angles invisible at planting.
Physical file movement and a central garden-state.md index (tracking stats, themes, ripeness, convergences, orphans, user patterns) make the system scalable: agent reads index first, avoiding all files every time. Result: open 40 half-thoughts without overwhelm; get dashboard of vital clusters and warnings instead.
Modular Rules and Skills Enforce Gardener Behavior
11 rule files divide into identity (user detection, config, patient voice), mechanics (file specs, cross-referencing via index, 5 ripeness criteria at 3/5 threshold), edges (no writing/research/pushing/deleting; composting at 14-day wilt/10-day orphan; pattern adaptation like shorter questions if ignored), and session integrity (sandbox to garden folder, end updates). This beats monolithic prompts: consistent, non-pushy agent adapts—e.g., prioritizes cross-referencing in planting bursts.
5 skills power 4 commands plus onboarding:
first-time-setup.md: Createsgarden/(inbox/seeds/ready/compost/garden-state.md), syncs Notion/Obsidian.greenhouse.md: Dashboard stats/themes/ripeness/convergences/orphans.plant.md: Single entry classifies new seed vs. signal; asks germination upfront ("What made you notice? Do you agree/resist?") for personal stake/tension.ripen.md: Checks 2+ criteria seeds, auto-moves at threshold, flags gaps (e.g., add tension).compost.md: Reviews candidates with context, cross-checks retired ideas.
Commands: "Show me the greenhouse", "Plant this", "Ripen", "Compost". Single plant command (killed separate signal) offloads sorting to agent; immediate germination (no queue) prevents backlog.
Ripeness Threshold and Germination Unlock Writable Angles
Harvest when 3/5 criteria hit: 1) Signal diversity (2+ sources vs. echo chamber); 2) Cluster size (2+ seed links); 3) Tension (unresolved question/contrarian); 4) Personal stake (your engagement); 5) Age (14+ days). Forces slow thinking: shower thought sits, gathers client convo/piece signals over 18 days, angle emerges. Germination questions embed stake/tension at plant—skippable but boosts ripeness. Agent flags unclear tension for input.
Use cases: Newsletter writers connect weekly plants into angles; consultants spot client patterns after 4 weeks; pros surface 14 expert angles. Killed abstractions (separate clusters/staging) simplify: filesystem mirrors metaphor, easing ops.
Build and Limits: Hands-Off Tending, Your Harvest Call
Paste article into Claude/Codex for step-by-step build (directory, rules, skills); premium gets pre-built RobotsOS download, onboard in <5 min via "Help me onboard". Agent learns patterns (plant frequency, theme dominance, response speed) without pushing—stays passive till summoned. Can't dictate next write (your judgment); reads external sources only on input, no auto-hunting (pair with WATSON for research). Trade-off: fewer decisions, better-informed via clean ready/ table.