Log

The idea was simple: scrape Reddit and other forums for founder complaints, extract pain points with AI, and surface product opportunities with citations. Better than manually browsing Reddit for 30 hours.

After 20 hours of building—10 data sources, keyword expansion, relevance scoring, clustering—the tool produced 9 pain points in 68 seconds at $0.50 per search. Asking Claude directly produces 50+ in 3 seconds at $0.01. The comparison wasn't close.

The fundamental mistake: trying to beat Claude at something it already knows. Claude's training data includes millions of Reddit threads, founder blogs, and startup discussions. No amount of real-time scraping could compete with that knowledge.

The most expensive lesson: validate the competitive premise before building. "Can web scraping beat Claude for founder problems?" should have been Day 1, not Day 3.

Major pivot. From "pre-build intelligence pipeline" to "The Toolshed"—a solo AI product incubator. One product per week, each exploring a new domain. The goal: learn whether building fast and selling cheap can compound into something durable.

The system that builds products becomes the product. Each week's build teaches something that makes the next build faster. The infrastructure compounds even when individual products don't.

First weekly build begins. Testing whether AI can extract actionable problem signals from startup communities. Validates the "problem discovery" piece—can we automate the manual Reddit research that founders do?

Target: ship something usable by Friday.

Scrapped evidence-scoring-layer approach. Deep dive into academic claim decomposition pipelines (FActScore, VeriScore, SAFE, DecMetrics). Validated that decomposition prompt quality is the bottleneck, not scoring algorithms.

Competitive landscape: Elicit, Scite, Consensus all academic-focused. Nobody building idea → build-spec pipeline. The evidence layer is commodity; structuring evidence into actionable build plans is the gap.

Cross-engine corroboration is a thin moat. The real bottleneck is idea → spec, not idea → evidence. Reframed as pre-build intelligence system.

Built full Launchpad UI at launchpad.theaiethos.com with dark theme. AI proposal generation: describe idea → get full app config. Proposal editor with tabs, build progress page with SSE streaming.

Fixed Clerk middleware issues, restructured CSS to app-level styling. The "describe and build" flow works—need to validate whether the generated specs are actually useful.

Built Blueprint (conversational project scoping), Bench (talent network), Retro (retrospectives). Full MVP loop: Scope → Staff → Deliver → Learn → Improve.

All three modules in ~3 hours across 2 sessions. Validates 1 MVP/week pace is achievable with AI-assisted development.

Launched standalone ethos site. Built /projects page with status badges, /log page for ongoing documentation. Established weekly log cadence.

Guildry Phase 0 complete: Clerk auth, Supabase with RLS, Claude API integration. AI creates clients via natural conversation. Pivoted from 17-PR waterfall to discovery-first approach—let AI conversations reveal what data matters.

Key learning: AI-assisted development is 10-20x faster than traditional estimates. Phase 0 took half a day vs 3 week estimate.

Previous iteration: Plinth as strategic research tool for consultants. Positioning defined but product scope kept expanding. Five build attempts with mixed results—consistent pattern of spec quality being the limiting factor.

The lesson that keeps repeating: better specs lead to better builds. The question is how to get better specs faster.