Prysai/Prysai-LLM-Playbook

What it solves

Newcomers face a confusing landscape of AI platform names — Codex, Claude Code, Agents, Skills — and don't know which one to learn first. This project reframes the starting point: understand what a language model can and cannot establish before choosing any tool. It also addresses a core failure mode of LLMs: a fluent, confident-sounding answer can still be wrong or silently add facts it was never given. The curriculum teaches one repeatable loop — define the task, choose a bounded action, inspect the result, keep evidence, state the limit — so readers verify model output before trusting it.

How it works

The default path is a five-unit LLM foundation core: explain what an LLM is and is not; write a small request with goal, context, limits, and output shape; identify omission, invention, forced ambiguity, and overconfidence; check and minimally repair an answer while stating one limit; and transfer the method to an unseen task. Each unit must leave a learner-authored artifact — a bounded task card, a checked result record, and a transfer attempt — rather than a copied prompt or polished model answer. A five-minute, no-setup prompt practice lets anyone see the invented-fact behavior in any chat model by asking it to rewrite a short message while keeping every fact identical, then checking whether it added details. The repository also maintains an evidence ledger that clearly separates what was observed (stable local engineering checks), what was captured but unscored (an output collection whose input hashes don't bind the prepared prompts), and what remains unknown (learner completion, transfer, productivity).

Who it's for

Newcomers to language models who want a tool-independent starting point and don't yet know whether to learn Codex, Claude Code, Agents, or Skills. No technical prerequisites: you don't need Codex, Git, a terminal, or a private file to begin — just any chat model for the practice exercise. It also serves readers who value honest status disclosure, since the project is explicitly a "candidate" with learner outcomes not yet measured, and it makes no IQ or productivity claims.

Highlights

• A five-unit foundation route built on a single inspectable loop: define, bound, inspect, keep evidence, state the limit
• A five-minute, no-setup prompt practice that reveals how fluent answers can fabricate facts
• Three learner-authored artifacts required per route: a bounded task card, a checked result record, and a transfer attempt
• An honest evidence ledger separating observed, captured-but-unscored, and unknown results — with no efficiency, IQ, or universal claims
• Eight registered language entry points and a public guided reading site; no technical setup required to start