Ask the library well

Who it is for: anyone connected to Nexorah Library. What you need: a connected AI (guide 2, 3 or 4) and an active plan.

How answers are shaped

The library answers in small pieces, so ask in steps instead of asking for "everything":

Steps

1. Give context first: say what you are building (type, platform, language). Run nl.py context in the project folder.
2. Search with plain words about the job, for example nl.py search "push notifications mobile".
3. Open the best result with nl.py card <slug>.
4. If a document is referenced, open it section by section: nl.py doc <path> --section 1, then the next number.
5. Check the verdict and the confidence label before you decide; read why when they surprise you.

Verdicts

• adopt — verified and safe for this kind of work; use it.
• trial — promising; try it on a small task first.
• assess — not enough evidence yet; research before relying on it.
• hold — risky or unverifiable; do not use it.

Your preferences

Your account can keep one short private note for your AI: your stack, languages, writing style, what it should always or never do, and the kinds of projects you work on. Edit it on your account page under My preferences (plain text, up to 8 KB), or ask your AI to change it. Your AI shows you the new text first and saves it only after you say yes.

Your AI reads the note at the start of each session, after the library rules; the library's safety and verification rules always come first. Only your own account can read it. It is never shown to other members, never added to the library and never sent as an experience report. Do not put passwords, keys or client secrets in it: the service refuses text that looks like one.

Example note:

stack: Flutter for apps, Cloudflare Workers + D1 for websites
languages: English and Spanish, right-to-left layouts supported from day one
style: short answers, result first
always: run the tests before saying done
never: copy other people's designs or code
projects: a booking website; a note-taking app

Reading a card's confidence

Every card answer carries a confidence block: a score from 0 to 100, a label (high, medium or low) and why, a few fixed sentences that explain the score. Search results and the sample cards of a Preview account show the label only; the score and the reasons come with the full card.

How the score is built (the defaults; the library may tune the numbers, the meaning of the labels does not change):

high starts at 70 and medium at 40. A hold card, and a card still queued for research, is capped below medium whatever the other signals say: a hold card is never more than low, and why then ends with capped at 39 (hold) or capped at 39 (queued).

What why tells you:

• verdict trial; 2 sources; tested in a project; 0 member reports; re-check due 2027-01-15 reads: promising, tried by the library, nobody else has reported yet. Try it on a small task first.
• re-check overdue means the card has passed its own re-verification date; see "In review" below.
• Most cards are low today, because most verdicts are still assess. A low label on an assess card is the normal state of a young card, not a warning; the warning is a low label on a card you expected to be adopt.
• The score comes from the library's own records, never from stars or popularity. Your accepted report can raise it, which is one reason to share (guide 6).

In review

A card also carries in_review. When it is true, the answer adds in_review_note: "Verified on <date>; re-check due". It means the library's own schedule says this card is due for another look: its re-check date has passed, or the librarian listed it for review. It does not mean the card is wrong. Treat the verdict as last verified on that date, look at the primary source for anything that changed since (version, licence, price), and prefer a high-confidence card for the same job when there is one. Search results carry in_review too, so your AI can prefer current cards before opening any. When you find a real change, flag it (guide 6).

The public library page shows one number for the whole library: the share of cards, hold excluded, verified in the last 90 days and still inside their re-check schedule. It is the same number your AI can read from the data; it tells you how fresh the library is as a whole.

Tips

• One topic per question. Short search words work best.
• Ask for the card when you need the exact command or licence; ask for a section when you need the method.
• Each lookup counts toward your daily limit (default 300); nl.py me does not.

Done means: you got a card with a verdict for your task and know whether to adopt, trial, assess or avoid it.

When the library has nothing on it

Sometimes a search returns no card. Your AI then says so plainly ("Not in the library yet") and researches the topic itself from primary sources: the official site, the licence text, the repository, the API documentation. It labels how sure it is (verified, rule of thumb, unverified) and never copies other people's text or code. With sharing on, it sends what it found as a finding report; the librarian verifies it, adds a card for everyone, and the report counts toward your Contributor reward. With sharing off, the research stays with you and nothing is sent, not even the topic.

Troubleshooting

• no_card or empty results: try other words, or say the library has nothing on it and research from primary sources.
• A document section ends abruptly: open the next section number.
• A card you trusted is suddenly in_review: its re-check date passed. Check the primary source for changes since the verified date; flag a real change (guide 6).

FAQ

• Can I download whole documents? No. Answers come one piece at a time.
• Can a hold card ever be high confidence? No. Hold and queued cards are capped below medium, whatever the owner sets the other weights to.
• Why did the same card change confidence between two days? A re-check date passed, a member report was accepted, or the library data was rebuilt with new sources.
• Does my AI see the formula? It sees score, label and the fixed why sentences; there is no free text, no adjective and no account name in them.