agora: Skill for Claude Code

Open-World Forge — break the loop by anchoring on the books, not on ourselves

The owner's complaint that created this skill: "I want to go farther than the same loop about the same research." The old idea-forge read the brain's own substrate (ideation/inputs = canon, beliefs, recent ideas) and so it recycled — coherence of what's-already-in-view manufactures confidence and the generator converges on its own attractor.

We have measured this exact failure: a generator that weights its own prior output with super-linear self-trust locks a permanent fraction of its founding bias (Lab 75db49, the strange-loop attractor — 50% locked at p=2; needs a ≥3–5% external-evidence anchor to recover). The books say the same thing three ways — Kahneman (WYSIATI: confidence comes from the coherence of what's in view, not its completeness), Gazzaniga (the interpreter confabulates a coherent story for outputs it didn't derive), Johnson (the adjacent possible: the reachable frontier is only what current memory makes visible).

So the cure is structural, not stylistic: keep the external-evidence fraction HIGH and search for novelty/absence, not coherence. This skill is built so that every run must stand on material from outside our own canon.

The four rules that make it open-world (do not skip)

1. External anchor every run. The primary substrate is the 36 books + cold vault domains, NOT our canon. The brain's self-knowledge is used for ONE thing only: to avoid repeating (the dedup/cold-domain set). Never let our own canon laws be the source of an idea.
2. Rotate the draw (divergence by construction). Each run reads a different random subset of books and a different set of cold domains, so two runs can't converge on the same combination. (No Math.random needed — vary by the date and what was drawn last, recorded in the note.)
3. Reward DISTANCE, not just feasibility. Score each idea on how far it sits from our current canon/recent-ideas. A safe, on-canon idea is the failure mode, not the goal. Surprise and "secrets" (Thiel) beat tidy coherence.
4. Open targets. There are NO fixed targets (the old 4 — os/agora/mcp_memory/realworld — were part of the loop). Generate first, anywhere, civilization-scale allowed; tag applicability after.