Let the library plan, review and check your work

Who it is for: anyone whose AI is connected to Nexorah Library and builds something: an app, a website, a reel, a Claude Code setup. What you need: a connected AI with an active plan (guide 2, 3 or 4). Preview accounts cannot use these three tools yet (they answer 402).

The idea in one paragraph

The library does not only answer questions. Before you start, it gives your AI a plan built from our tested guide for that kind of project. After your AI has built something, it gives a review checklist made from the same rules our own reviewer uses. Before your AI says "done", it runs a done-check that asks for evidence. None of this uses a model on our side: the library assembles structure from its verified guides and your own AI does the thinking. Each of the three calls counts as one question of your daily quota, like a search.

When to use which tool

Plan

Give the task in a sentence or two, plus optional signals such as the stack or languages:

library_plan: task = "Flutter app for a bakery with three languages", signals = ["flutter", "dart"]

What comes back:

• project_type (app, website, social-media or claude-code) and guide, the library document the plan follows.
• steps: one step per stage of that guide, in order, each with a title, a guard (what must be true before the step counts as done) and read: the document path and section number. The plan gives pointers, not text. Ask your AI to open each step's section with library_doc before working on it; the full instructions, numbers and links live there.
• cards: the tools to use in that step, with their verdict and one line on when they fit; only tools the library has verified or is trialling.
• ask_first: what only the person can answer (accounts, data location, payment). Ask these before building.
• more_stages: when a guide has more stages than fit in eight steps, the rest are listed here with their read pointer; they are part of the work too.
• rules_reminder: the numbers of the library rules that matter for this task.

If the library has no guide for the task, the answer is the honest one from guide 5: "Not in the library yet", with the instruction to research primary sources and send a finding. Nothing is invented.

Review

Tell the library what kind of work it is and what was built, and add the facts you have:

library_review: kind = "ui", summary = "Account page with a connections table, dark and light", facts = { widths_captured: [1440, 390], tests_before: 212, tests_after: 220 }

What comes back:

• checklist: numbered items, each with verify (how to check it: which width, which transition, which command) and source (the library section the item comes from, readable with library_doc).
• acceptance: the three rules our reviewer applies: every item or reject; compare with the previous version and with the best accepted one; report defects as a numbered list of defect, place and fix.
• missing_facts: the facts the checklist expects and you did not give, such as a mobile capture or the test count before the change.
• truncated, when the list was cut to stay small; the dropped items are still in the source sections.

The library does not judge the work. Your AI does the review, from its own screenshots and runs, and the checklist keeps it honest.

Done-check

List the claims and the evidence:

library_done_check: claims = ["account page done", "tests green"], evidence = { tests_before: 212, tests_after: 220, journeys_passed: 17, journeys_total: 17, screens: [1440, 390], remaining: ["none"], committed: true, nothing_running: true }

What comes back:

• verdict: ready only when nothing is missing; otherwise not_ready.
• missing: the numbered gaps, from fixed rules: a remaining-work list must exist (even "none"), tests after must not be fewer than before, journeys complete, both widths for interface work, committed, nothing left running.
• unproven: claims that have no evidence behind them.
• template: a ready "Remaining work" section to paste into the report.
• note: "Checklist verdict from self-reported evidence; not a certificate." The library cannot see your tests or screenshots; it checks that you reported them. A green done-check with false numbers is your AI's mistake, not proof.

From the terminal (Claude Code plugin 1.3.0 and later)

The same three calls exist as commands of the plugin script, so a shell step or a hook can use them: nl.py plan "<task>", nl.py review <kind> --summary-file F [--facts-file F], nl.py done --claims-file F --evidence-file F (exit code 2 when not ready), and nl.py report-status <id> for a report you sent. Add --json for the raw answer. The script sends technology names only, never file names or paths, and cleans your text of keys and emails before it leaves.

Steps

1. New task: run library_plan, ask the person the ask_first questions, then open each step's section with library_doc as you reach it.
2. Build step by step; the step's guard is the test you must pass before moving on.
3. Before delivery: run library_review for the kind of work, do every item from your own screenshots and runs, and fill the facts it says are missing.
4. Before saying done: run library_done_check with real numbers; fix what missing lists; paste the template into your report.
5. If a step taught you something new, send a report (guide 6).

Done means: your AI's delivery report ends with a remaining-work section that came from a ready done-check, and every review item was checked from evidence, not from memory.

Troubleshooting

• plan_guide_missing (404): the guide exists but has no English mirror yet; tell us through a report and use library_doc on the guide directly.
• review_source_missing (404): a configured section of that kind has no English checklist yet; the other kinds still work.
• The plan has only one step (Claude Code tasks today): that guide has one section; read it with library_doc.
• 402 on all three: your account has no active plan or is in Preview.

FAQ

• Does the library read my code or screenshots? No. It never receives files; it receives your summary, facts and claims, and answers with structure.
• Why eight steps at most? Answers stay small on purpose (under 4 KB) so every AI client can read them; the rest of a long guide is in more_stages and in the sections themselves.
• Can a done-check be wrong? Yes, if the evidence you gave is wrong. It is a checklist, not an audit.
• Does this cost extra? No model runs on our side; each call is one question of your plan.