Capability · Wedge · Ideas
The idea kiln.
Type a paragraph. Six stages later, you have three validated micro-SaaS hypotheses that survived a hard kill filter and a Market Skeptic. Every survivor is anchored on a capability wedge less than 90 days old — competitors haven't shipped against it yet.
Most AI idea generators are slot machines.
Pull the lever, get fifty plausible-sounding hypotheses, half of which die in week one because they were built on yesterday's capabilities or land in a red ocean. HypoKiln is the opposite of a slot machine.
Brainstormer
- "AI-powered note app" is an accepted output
- No source for the capability claim
- No competitor check, no pricing reality
- Looks confident — survives by sounding plausible
- You find out it's dead at month six
HypoKiln
- Refuses ideas without a primary-source URL for the wedge
- Wedge must be ≤90 days old or it gets archived
- 13 named hard-kills + 5-persona stress test before approval
- Three deterministic structural audits, not vibes
- You find out it's dead in 30 minutes
How it works
Six stages. One human gate.
No deploy, no build, no bullshit — the kiln stops at G1, where the highest-leverage decision lives. Three stages run under a critique loop: author drafts → deterministic gate runs → if FAIL, a critic agent writes structured feedback → author iterates.
↻ marks stages wrapped in the critique loop. The author can't pass a stage by sounding confident — only by passing the deterministic gate.
The capability radar
No wedge → no idea.
factory/00-radar/capability-wedges.md is a canonical log of AI/LLM/media releases from the last 90 days. Stage 4's Kill Filter hard-kills any hypothesis whose capability_wedge.id isn't on the active list, isn't fresh, or doesn't have a primary-source URL.
"AI-powered" is rejected.
"Batch tool-calls drop scan cost 10× — Anthropic, 2026-04-12" is accepted.
One-line install
Pick the row that matches your machine.
What lands on disk
Markdown and JSON. Nothing hidden in a database.
Every run drops a self-contained folder under products/<slug>/. You can cat your way through any idea, and every artifact is the input for the next stage. No DB, no migrations, no schema drift.
products/<slug>/ ├── research/ │ ├── trend-radar.md 10+ dated signals with URLs, 3+ distinct sources │ ├── round-001.json 10+ hypotheses, each with capability_wedge + viral_mechanic │ ├── market-snapshot.md why-now + recent-date evidence │ ├── competitor-analysis.md 3+ named competitors with URLs and moat-depth scores │ └── pricing-research.md 2+ price tokens, willingness-to-pay quotes ├── spec/ │ ├── decisions.md append-only cross-stage memory log │ ├── architecture.md form_factor + archetype + wow_moment + viral_mechanic │ ├── gate-1-preflight.md 10-question alarm-count checklist │ └── gate-1-approval.md signed (or rejected) verdict └── .critique-log/ per-iteration transcripts + critic feedback
Bring your own coding CLI
Run on the subscription you already pay for.
HypoKiln never talks to provider APIs directly. Each stage spawns a logged-in coding CLI as a subprocess; auth lives in the CLI's session file. HypoKiln cannot read your token even if it wanted to.
OpenAI
- Login:
codex login - Plan: ChatGPT Plus / Pro / Team
Anthropic
- Login:
claude→/login - Plan: Claude.ai Pro / Team
- Login:
gemini auth - Plan: Google AI Studio
Most teams get the which-idea decision wrong.
Then they burn the next six months building something that was always going to die. You don't need a kiln for the build. You need a kiln for the decision.