The 649-entry guide is Chinese and mainland-specific
HowToLiveBetter is an open book organized as a ranked set of practical choices. Its 649 entries cover health, emergencies, money, work, family, travel, online services, and legal risk. Each entry states a cost, a plain-language outcome, an evidence grade, a source, and any caveat. The content is Simplified Chinese, and sections on law, insurance, benefits, employment, and permits follow mainland China's rules.
English readers get a clear warning rather than a false promise. The main repository has no English README, and its language links point to a separate fork maintained by another contributor. That fork includes English, Russian, Spanish, and Portuguese, but the Chinese README says outside translations may lag and treats the Chinese original as authoritative. For changing laws or benefit amounts, a delayed translation can change the decision rather than merely the wording.
Search filters expose cost, evidence, and source
The online page can filter by chapter, evidence grade, money, time, and effort. EPUB, PDF, and a single-file offline HTML version are attached to a rolling release. The static site reads the book files directly, so there is no account, database, or backend to operate. The content license is CC BY 4.0, while tools, skills, workflows, and the search page use MIT.
This is more useful than a loose advice list because readers can move from a recommendation to its cited paper or official document. The format still contains editorial judgment. Choosing a cost category, assigning evidence grade A, or ranking one action above another requires decisions about population, baseline risk, and whether a legal requirement counts as a measurable benefit. A visible citation lets you audit that judgment. It does not make the judgment disappear.
What happened when we ran it
Our sandbox cloned commit dc74035 and worked inside tools/epub. Installation took 6 seconds, added 1 package, and occupied 1 MB. The EPUB build completed in 5 seconds. We used an unprivileged Debian container with Node 22, 3 CPUs, 8 GB of RAM, and no secrets. Npm audit reported 0 known vulnerabilities: 0 critical, high, moderate, or low.
No test script or target existed, so the test step was skipped. The repository had 177 files, about 1,102 lines of source code, and a 3.8 MB checkout. Our scan found 1 CI workflow, no Dockerfile, and no tests directory. Those figures describe the publishing tools, not the much larger prose corpus, and a successful EPUB does not verify a health claim or legal interpretation.
Most readers should not build anything. The hosted search page and release downloads already cover online, offline, print, and e-reader use. A local static server is enough for the repository page. Rebuilding the PDF is the heavier branch because the README calls for Pandoc 3.1 or newer and Typst 0.13 or newer. The measured 5-second step covered the EPUB command only.
Issue 61 found nine repeatable reasoning problems
A review opened on October 1 reported nine high-confidence problems found by at least two independent readers. Examples included relative risk without baseline risk, legal obligations presented as benefits, and evidence grades that did not match the strength of the conclusion. The issue was closed the same day, and the repository received another push. That response is a good maintenance signal, while the report itself is a warning against reading grades as settled truth.
The distinction matters most in high-stakes chapters. A population study may not describe your medical risk. A national average may not equal the payment in your city. A legal deadline can have exceptions that a short entry omits. Open the cited source, check its date and jurisdiction, and get professional advice when the cost of being wrong is high. The guide's traceability makes that check possible.
The AI skill retrieves entries before it answers
The repository includes one skill shared by Claude Code and Codex. Its rule is narrow: search the local book first, rank the relevant entries using the book's method, name the section and item, and say when the source does not answer. That is a better contract than asking a general model to remember 649 entries, especially when the content changed after the model's training data.
The skill does not cure source errors or translation drift. It can retrieve an outdated number faithfully, and a local clone can fall behind the October 1 source. Keep the repository fresh, require section references in every answer, and open the underlying citation before acting on medical or legal advice. HowToLiveBetter is worth using because it leaves an evidence trail. Its value ends where the reader stops following that trail.