HowToLiveBetter Puts a Price and an Evidence Grade on Every Life Decision
The result is a transferable pattern for turning a body of practical knowledge into a queryable, auditable dataset: every claim carries a primary source, every judgment is labeled as a judgment, and AI assistants can answer from the text with citations. The China-specific legal and policy entries won't transfer directly, but the structure and tooling transfer cleanly.
HowToLiveBetter, created in September 2026 and already at 67,897 stars, is a 34-topic collection of 767 suggestions for living longer, spending less, avoiding legal trouble, and handling practical paperwork. Each entry follows a fixed format: cost in money, time, energy, or willpower; a plain-language statement of what changes; the underlying numbers; an evidence grade; a primary source; and the boundaries of who the advice applies to. Payoffs are kept in four non-interchangeable categories: lifespan, time/energy, money, and personal freedom, with no conversion between them.
HowToLiveBetter treats its content as software. Markdown files are the single source of truth; a static search page reads them directly, and build scripts regenerate an offline single-file HTML, a PDF, an EPUB, and an Anki deck on every content change. A companion skill lets Claude Code and Codex pull relevant entries, rank them by the book's cost-benefit method, and cite section and item numbers. Verification logs are public, and errata go through GitHub Issues.
Evidence and value are deliberately kept apart. A/B/C grades only say whether a number is checkable: 495 A-grade items come from meta-analyses, large cohorts, or RCTs; 220 are B-grade; 52 are C-grade. A separate cost-effectiveness tier, labeled as the author's own judgment and therefore only C-grade itself, ranks 123 items as extremely high value, 344 as high, and 300 as average. Legal and policy content follows mainland China rules as of writing.
Separating 'is the number trustworthy' from 'is this worth doing' is the book's strongest editorial move. Evidence grade A does not make an item worth doing, and the cost-effectiveness tier is explicitly labeled a personal judgment.
Keeping lifespan, money, time, and legal risk in separate accounts avoids the false precision of converting everything into one metric. The result is a recommendation list that can honestly say a 12% mortality reduction and 500 yuan a year are not on the same scale.
Treating the book as a build pipeline is what makes a 767-item knowledge base actually maintainable. Single-source Markdown, auto-generated exports, public verification logs, and issue-tracked errata turn a document into something a community can keep correcting.
The 512 KB GitHub rendering limit forced the content into 34 files; that constraint is a reminder that platform limits often shape repository architecture in ways a design document never would.
The AI skill's value depends on the underlying text being structured the way it is. Because every entry has fixed fields and cross-references, an agent can answer with section numbers instead of hallucinating, which is not true for most unstructured documentation.