Give the assistant a small task with a stopping point
An assistant can help organize your existing material and administer an authorized rehearsal. You still need to check the questions and approve publication. Participants answer formal exams under their own identities; management automation does not grant permission to sit an exam for someone else.
This walkthrough uses the English pilot: twelve fictional questions and two participant rows. Download its rules, source JSON, prepared candidates, expected answers and participant JSON. Read the source before giving the assistant an instruction.
The sample's four True/False statements use English options in single-choice questions, alongside four other single-choice and four multiple-choice items. It does not exercise the separate judgment type. Use a separate pilot for other required formats. File verification and the documented steps below are not evidence that your organization has already completed an exam.
Connect with the intended organization permissions
Sign in normally to Learnweft and choose an organization you are allowed to administer with sufficient active service. Record its actual organization UUID and verify its name. A member's short organization code is not that UUID. The owner and two ordinary rehearsal members need at least three available member places.
In Personal space, open AI assistant and expand its organization-management configuration. Use the current configuration supplied there for your client. This walkthrough follows the existing Codex MCP path. The endpoint is:
https://apis.lihuanyu.com/question-bank/mcp
Complete the normal login flow from the provided configuration, then return to AI assistant, refresh the connection and explicitly authorize the rehearsal organization and bank scope. For a configuration named zhixu, the client's login command is:
codex mcp login zhixu
Check the resulting permissions and expiry. Creating a new bank requires authorization that permits a new bank, rather than access only to an unrelated existing bank. A personal connection does not automatically acquire organization-management access. See the official Codex MCP documentation for client configuration; the current Learnweft page defines this service's scopes.
Review the source locally before any write
Start by asking the assistant to compare the neutral English source and expected answers. The prepared candidate JSON is for structured import, while the DOCX is for the normal upload form. Do not mix those entry points.
Read rules.md, source-questions.json and expected.json locally.
Do not call Learnweft write tools yet. List the twelve normal questions,
options, correct option keys and explanations. Compare every answer
with the fictional rules, and flag any disagreement for my review.
Do not include G013 or G014 from the separate boundary sample.
Do not invent missing answers or claim that a file hash proves correctness.
Confirm the intended answers yourself. The JSON carries source-file references and a SHA-256 digest. If the source changes, rebuild the candidates and review again. A digest binds bytes; it does not approve an answer or authorize sending private documents to an external model. This rehearsal uses public synthetic material.
Preview a bank and import, then confirm publication
Replace the placeholder below with the organization UUID you checked. Have the assistant read the current tool schemas instead of constructing parameters from this article as if it were an API definition.
Operate only in organization REHEARSAL_ORGANIZATION_UUID.
Read back its name and current authorization. Preview a new bank named
"English pilot bank" for practice and team exams, then wait for my approval.
After approval and creation, use its real returned ID to preview
create_prepared_import with training-candidates.json and all twelve
normal candidates. This public synthetic material is authorized for
storage and processing, so rightsConfirmed is true.
Use a stable idempotency key for each logical write. Do not publish yet.
After the import, have the assistant read back all candidates with their actual row versions and issues. Confirm or repair them individually. Do not accept the bank solely because the tool reported twelve rows.
Show each candidate beside its source question. Accept only the items
I explicitly confirm. Read current versions before every change.
After review, preview publish_import and wait for my approval.
Then read the current bank revision, preview publish_bank_version,
show the count and target, and wait for separate approval.
Read back the published version and its twelve usable questions.
Previews and confirmation credentials bind the operation, inputs and resource version. If the target or content changes, read and preview again. Candidate creation, reviewed content and a published version are distinct states.
Set up two participants and a future exam
The participant JSON contains ordinary fictional employees without contact details or passwords. If its employee numbers already exist, choose unused rehearsal identifiers and use the same file throughout preview and confirmation. Do not modify an existing employee to make the demonstration fit.
Preview only training-members.json for this organization. Show employee
numbers, job-position matches, row count and remaining member capacity.
After I confirm, execute preflight and read the pending batch.
Preview final member confirmation and wait again before importing.
Do not set passwords, retrieve activation codes, change roles or touch
other members. Return the actual IDs and states of these two members.
An administrator and the participants complete the normal Web password-setup and sign-in flow privately. Imported or pending rows are not proof that a participant can enter an exam. Check access before publishing. For a small Web-only alternative, add the same two fictional people manually through member management.
Ask the assistant to prepare an exam with the published twelve questions, those two verified participants, ten points per question and 120 points total. A passing score of 72, one initial attempt and a short duration are rehearsal choices. Show the intended retake and score-selection policy before choosing it.
Have it propose a future availability window with an explicit timezone and UTC offset. Check the form's timezone and the exact instants rather than assuming an English interface changes the clock. Preview the complete draft, then separately preview publication and wait for your approval. The plan's remaining service must cover the rehearsal and any retake.
Read actual results and close the authorization
Have each participant answer and submit in Web. The assistant may then read authorized management results and export them; it must not create fictional submissions or alter scores. Follow pagination until the full roster is reconciled. Compare participant IDs, the published version, answers actually submitted, scores and the report's selected attempt.
After a timeout, query the management request receipt using the original operation, organization and idempotency key before retrying. Do not create a new key just because a response was missing. On a version conflict or denied scope, reread the state and resolve the target or authorization rather than trying a different organization.
Finally, narrow the connection to the bank you need or revoke it in AI assistant. Keep credentials outside articles, prompts and shared files. API-specific REST tokens do not replace MCP OAuth. The verified CLI management package is available through a controlled channel; this guide does not require an unpublished public installation command.
For the full practice and retake checks, return to the English pilot guide. For purchasing, check the actual currency, payment eligibility and service conditions on the pricing page. Assistance with preparation is useful only when the reviewed content and actual participant results remain verifiable.