BUILD WITH MODELS

Skills, agents, and the repo

Drop-in specs for coding agents. MIT. Original text only — no book pages or figures.

skills/many-model-selector

pick: 3-5 families, max diversity
check: [tails, incentives, contagion]
output: lattice + where they disagree
hedge: at the disagreement

skills/redcape-reasoner

for job in [R,E,D,C,A,P,E]:
  apply 1 model, note failure mode
stop when 2 models disagree
that disagreement = the insight

skills/distribution-checker

if loglog-straight: power-law, never average
elif bell: normal, predict intervals
elif bell-after-log: lognormal, work in logs
else: mixture, segment first

agent-skills/apply-three-models

for m in 3 distinct families:
  conditions, prediction, breaker
end: one hedged action
     + the reversal trigger

Try it: paste into any agent

Apply 3 models to opioid approvals: Bandit + Markov + Dynamics. Hedge.

Expected: trial arms, a transition matrix, and a loop map. The full specs live in skills/ and agent-skills/ next to the app, and the catalog is one call away at /api/models.

datamodellib/
  app/  library/ model/[id]/ playbooks/ mentor/ skills/ api-docs/
  app/@modal/(.)model/[id]/   <- card clicks open modals
  app/api/models/             <- read-only JSON catalog
  app/api/mentor/             <- mentor (Groq → Mistral → Gemini → OpenAI), off by default
  lib/models.ts               <- 38 original summaries in 10 families
  lib/mentor/                 <- providers, tools (web search, catalog), agent loop
  skills/ agent-skills/       <- agent drop-in specs