wildcard
globalanomalyindex/wildcard
an experimental claude code skill: a synthetic default-mode network that seeds non-linear creative thinking with real entropy
Instalação
npx convez-ai globalanomalyindex/wildcard1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
wildcard
an experimental skill for claude code that, at random, summons a wildcard from outside your problem to seed fresh, non-linear thinking on whatever you are working on. the draw lands in one of two modes: a hyper-specific niche specialist from a completely unrelated field, or a general concept to connect against. in doing so, the token pool is seeded with unorthodox directions for standard chain-of-thought systems to work with, influencing more creative thinking, problem-solving, and idea generation in large language models.
an llm's chain of thought has no default mode: it is relentlessly task-locked and never wanders, so the human "pop-in", the remote associate surfacing from the subconscious, cannot happen; the architecture forbids it. wildcard externalizes that missing faculty: a synthetic default-mode network that injects the remote associate the on-task reasoner would never wander to.
the wildcard is a seed, not a subject. you do not force your problem to map onto it; you let it steer your thinking somewhere new and bring back the genuinely good idea it seeds, stated in your own terms. it never invents a connection, and it abstains honestly when the draw leads nowhere good.
live site: https://globalanomalyindex.github.io/wildcard/
and yes, i tested whether it actually works: a pre-registered, blind, three-arm study (the model's own "random" picks collapse and hug the problem; the external draw does not), an honest result that went against me at first, a diagnosis, a fix, and a re-test on fresh problems that confirmed the fix out of sample. the whole write-up, with every number regenerable, is the case study: https://globalanomalyindex.github.io/wildcard/case-study/
install
in any claude code session, two commands:
/plugin marketplace add globalanomalyindex/wildcard
/plugin install wildcard@globalanomalyindex
that's it. the plugin pulls the skill, its draw.sh, and both pools; updates arrive when you run
/plugin marketplace update. then, in any session:
/wildcard
prefer a manual install? drop the skill folder into your personal skills directory:
git clone https://github.com/globalanomalyindex/wildcard
ln -s "$(pwd)/wildcard/plugin" ~/.claude/skills/wildcard
how it works
- detect, then freeze - distill a structural sketch of your project (moving parts, flows, tensions) from its files and the live conversation, and commit to it before drawing. structure, not stack: it works on code, writing, design, or research alike.
- draw, outside the model -
draw.shfirst rolls a mode, then rejection-samples one of 378 niche disciplines or 461 concepts with real os entropy, so every choice is exactly equiprobable and comes from outside the model's own distribution. (--seed Nfor reproducible draws; the live site reproduces any seed in your terminal.) - become it - inhabit the wildcard and think from it, not about it. that inhabiting is the conditioning that does the seeding.
- harvest what the seed surfaces - keep only what passes the removability test: delete the sentence that names the wildcard, and an executable move in your own words must still stand, one the plainest reading of the problem would not already produce. structure-mapping (gentner: relations, not nouns) is the sharpest tool here and the check against decoration, not a cage.
- present seeds - a few tight, optional seeds, each ending in a concrete move with a falsifiable
specific, with an offer to pull one thread further. optionally banked to
.wildcard/seedbank.mdin your project.
guarantees, each enforced by a mechanism rather than a hope: no fabrication (only genuine, usable ideas; honest abstention is honorable), no decoration (the removability test rejects a visible-but-undischarged analogy even when the mapping is real), no derailment (seeds are additive and optional, never prescriptive), no favoritism (the dice live in a shell script, not in the model's associations), and inspiration only (a drawn title or concept is a pointer, never content to reproduce).
layout
.claude-plugin/marketplace.json # marketplace catalog (one plugin)
plugin/ # the installable plugin
.claude-plugin/plugin.json # plugin manifest
SKILL.md # orchestration the model follows
scripts/draw.sh # the dice outside the model (mode roll + rejection-sampled entropy)
scripts/audit_domains.sh # breadth/format gate for the discipline map
scripts/screen_concepts.sh # logged safety filter for the concept pool
scripts/audit_concepts.sh # format/tier/denylist gate for the concept pool
references/domains.txt # 378 niche disciplines, axis-tagged for coverage auditing
references/concepts.txt # 461 safety-screened concepts (wikipedia vital articles snapshot)
references/specializing.md # persona growth, anti-mode-collapse
references/structure-mapping.md # the honesty bar (servant, not gate)
references/connecting.md # the loose-to-genuine refinement protocol for concept mode
tests/run_all.sh # full suite (draw, audits, real-pool diversity, experiment libs)
site/ # the landing page (globalanomalyindex.github.io/wildcard)
experiment/ # the pre-registered study + the seeding fix re-test (frozen data)
docs/ # case study, methods colophon, design specs + plans
verify
bash tests/run_all.sh # expect: ALL GREEN
the suite proves the load-bearing properties: seeded draws are deterministic and reproduce the
website's live draw byte-for-byte; entropy draws are rejection-sampled to exact uniformity; both
pools pass their breadth and safety audits; and many draws over the real pools stay widely varied
with no dominance. ci gates the live-site deploy on this same suite, so the page can never ship a
claim the mechanism fails. the experiment's numbers regenerate too: node scripts/analyze_experiment.mjs experiment and node scripts/analyze_retest.mjs experiment/v2.