Keep user-confirmed information separate from AI interpretations and external facts. Retain the source and date of material information. When a correction conflicts with an earlier answer, clarify what has changed instead of silently choosing whichever value produces a convenient recommendation.

Updating the user profile should not rewrite a historical application, a provider response or an offer. Keep that history and obtain any new approval required for a revised action.

Interest, preference or personality activities may help a conversation when their purpose and limitations are clear. They should not be presented as a diagnosis, proof of ability or a fixed limit on what you can become. Do not infer sensitive characteristics or reject a route because a personality label appears to fit another one better.

For a school student, identify near-term subject, study and application decisions without demanding a lifelong career choice. For a current learner, compare actual learning with intended work. For a jobseeker, investigate the barrier before prescribing training. For a career changer or returning worker, preserve useful experience and examine what needs updating.

The person: goal, constraints, preferences, confirmed history, evidence and unresolved questions. The opportunity: identifiable provider or employer, specific offering or vacancy, relevant requirements and its current status.

The relationship: why a learning activity, task, tool, credential or experience supports a particular route. The action: the approved request, information shared, recipient, response and next owner. The result: what actually happened, when and how it was confirmed.

Record the source identity and location; the relevant statement or evidence; publication or effective date when available; retrieval and verification dates; geography and population; the course, offering, role or task it refers to; and any known limitation.

Keep provider claims, user reports, measured observations and AI interpretations distinguishable. A recent retrieval date does not prove that an old course page or job advertisement is current. An absent date must remain absent, not be invented.

Distinguish a provider from its campuses and offerings, a qualification from a specific delivery, and a vacancy from repeated advertisements. Retain meaningful differences rather than merging records because their titles are similar.

When sources conflict, preserve the conflicting statements and seek confirmation from the appropriate authority. Explain which point remains unresolved and how it affects the recommendation. Do not silently select the cheaper fee, easier requirement or more attractive outcome.

Use a review schedule appropriate to the field and the consequence of error. Vacancies, offer deadlines, fees, funding and intake availability can need a new check before action. Stable descriptive material may be reviewed differently.

Flag stale, withdrawn and superseded information. Recheck decision-critical details before sending a request or relying on an offer. This method does not claim that every record is refreshed in real time.

Use the observation period, geography, source coverage and occupational level when describing demand. Deduplicate appropriately and distinguish broad occupation trends from specific entry-level opportunities.

JSA notes that its Internet Vacancy Index does not cover all recruitment channels and that advertisements do not map one-to-one to vacancies. A small collection of advertisements must not be presented as the whole labour market or a personal probability of employment. [7]

Keep the task, underlying workflow, product name, relevant version or configuration, expected proficiency and working context separate. Do not infer that a tool is obsolete because it is less visible in a small sample, or that a new brand requires a whole new qualification.

For a replacement or trend claim, identify the dated evidence and the setting in which the change applies. JSA describes the effects of generative AI as varying by occupation, industry and context; that is a reason to investigate task-level change, not predict the future of every individual role. [1]

Say what is unknown, identify the source or person who could resolve it, and explain whether it blocks action or only limits confidence. The system may offer a conditional comparison or pause a consequential action.

No result within the sources searched does not mean that no opportunity exists. A limited catalogue should be described as limited. Do not use a fluent answer to conceal insufficient evidence.

Retain a record of the correction and its basis. Revisit affected comparisons, scores and next steps. Where an external request has already been sent, identify whether a corrected message or renewed approval is needed.

Keep enough version history to understand why a recommendation was made at the time, subject to the applicable privacy and retention arrangements. Never alter a historical offer to make the record match a later assumption.

Start with the goal and constraints, find candidate routes within the documented coverage, check mandatory conditions, then compare meaningful alternatives against the same criteria. Preserve material trade-offs rather than presenting one option as universally best.

The user should be able to understand exclusions, challenge assumptions and consider another route. The number of options shown should serve the decision, not create a false impression that every possible pathway has been simulated.

Before a numeric fit score is used, define its dimensions, normalisation, weights, treatment of missing data, relevant comparison set and method version. Show the reasons behind it and keep eligibility and evidence confidence separate.

A fit score is not an admission probability, a hiring probability or a judgement of a person's worth. No validated production scoring formula or performance result is published with this edition. The examples therefore do not invent match percentages.

A missing prerequisite, unclear eligibility, unavailable place, unconfirmed funding or non-negotiable timetable conflict must remain visible. There is no blanket percentage tolerance for mandatory requirements; any alternative entry must have its own valid basis.

Unknown capability must not automatically count as no capability. Missing information may reduce confidence or prevent a score, rather than lower the person's apparent suitability.

A text-similarity score may help find a course, role or relevant passage. A conversation-routing score may help choose the next screen or tool. Neither, by itself, establishes that a pathway is suitable or that a person meets the requirements.

The method must make the substantive comparison against the person's goal, circumstances and evidence after retrieval, not simply rename a technical similarity score as a career-fit score.

Pause when a consequential decision depends on contradictory, missing or insufficiently current information. Explain the unresolved question and the next verification step. Escalate specialist questions to an appropriate person rather than extending AI authority beyond the evidence.

A user may still compare possibilities while a fact is unknown, but a conditional possibility must not be represented as confirmed eligibility, a secured place or an authorised action.

“Use a CRM” might mean entering accurate records, retrieving information, handling a workflow or administering the system. Check the actual task, expected independence, context and product-specific detail before deciding what a beginner must learn.

Likewise, a course page that says “business software” does not establish either that a specific CRM is taught or that it is absent. Ask about access, practice, feedback and assessment. Keep the course-content question separate from the person's capability.

A project demonstrates work produced under particular conditions. Identify the learner's contribution, assistance received, task difficulty, reviewer and standard. Distinguish an AI-generated draft, a guided exercise and independent workplace performance.

Use fictional, public or specifically approved data. Do not put private customer records into an exercise without the required permissions and safeguards. A learning activity must not quietly become unpaid production work.

You make the choice about your pathway and the actions you authorise. An employer makes its hiring decision. A provider determines its admissions, assessment and recognition outcomes. A regulator or authorised specialist handles matters within its authority.

Human review does not create authority that the reviewer does not hold. Refer legal, migration, financial, clinical or other specialist questions to an appropriately qualified service rather than presenting general pathway guidance as a formal determination.

With the appropriate permission, pass on the decision being considered, confirmed context, options compared, evidence, unresolved questions and relevant action history. Avoid requiring the person to repeat sensitive information unnecessarily.

Explain who is helping, what they can resolve, any cost or service boundary and the next step. Do not imply that a handover, response time or ongoing counselling arrangement exists unless it has been confirmed.

Keep conversation, exploration, saving, sharing, requesting, applying and accepting separate. Explain the destination, information and consequences before an external action. Obtain the permission appropriate to the action and the person's circumstances.

Let the person pause, decline or choose an alternative. Do not contact a provider or employer just because the person clicked a comparison or mentioned an interest.

Use understandable language, manageable steps and formats that suit the person. Avoid making a long questionnaire the only route to help. Check relevant age, service and permission requirements rather than assuming the same process applies to everyone.

UNESCO's guidance emphasises human-centred, age-appropriate educational use of generative AI, with attention to privacy and validation. It does not certify Yuzee or determine every local requirement. [4]

Collect and share what the task needs, not every detail that might be useful someday. Explain the purpose, recipient, access, retention and applicable correction or deletion process in the actual service. Do not infer sensitive traits from conversation or use them to silently narrow someone's future.

This HTML page has no account connection, analytics or live AI. Its brief tool uses page memory only. The privacy behaviour of the wider Yuzee platform must be described in its own applicable notices and controls.

Make paid promotion and relevant commercial relationships visible where they affect access, presentation or the transaction. A commercial partnership is not evidence of personal suitability. Explain catalogue limits and retain genuine alternatives, including routes that do not require a paid course.

Do not optimise the person's recommendation solely for enrolment, commission or engagement. A worthwhile next step may be work, recognition, support, another provider or no immediate purchase.

Make clear which answers are general guidance and which decisions belong to a provider, employer, regulator or authorised specialist. Explain when further verification or human help is necessary and what can be done while that question remains open.

Do not present generated skill labels as a hidden assessment, a diagnosis or a sufficient basis for automated rejection. A person should be able to correct information and challenge an important interpretation.

Preserve enough of the source, method version, decision context and action history to explain a result appropriately. Identify and repair errors in affected recommendations, and communicate material corrections where necessary.

The OECD AI Principles include transparency, accountability and traceability across data, processes and decisions. This informs the standard; actual implementation still requires evidence and evaluation. [5]

State the evaluated service and method version, period, population, sample size, denominator, measurement definition, follow-up and limitations. Distinguish user feedback, expert review, observational results and comparative evidence.

Use an appropriate baseline or comparison before attributing an improvement to Yuzee. Track adverse outcomes and material errors, not only successful cases. Test performance across relevant situations and access needs without treating one average as sufficient proof.

No production accuracy rate, hiring improvement, learning gain or return-on-investment result is asserted on this page.

Before a person proceeds, the service should identify the supported request type and market; its availability or pilot limits; the human support included; the responsibilities for checking, sending and following up; costs; information-sharing conditions; and what happens when the request cannot be fulfilled.

A capability should be described as available, limited pilot, prototype or planned using a dated, owned verification record. This edition is not connected to that deployment register and does not assign live status to unverified platform capabilities.

It does not establish a complete market catalogue, validated scoring formula, production test pass rate, provider partnership, guaranteed offer, automatic funding entitlement or measured learning or employment improvement.

The principles and examples provide an explanation of intended behaviour, not independent evidence that each standard has been implemented. A measured claim requires the relevant product, data and evaluation records.