The same words can lead to different needs
“I want someone to play table tennis with” and “I want someone to coach me at table tennis” share a sport but describe different intentions. One is a social invitation; the other could be paid work. A system that asks for a budget in both cases has missed the important distinction. Neeed’s interview is being built to clarify meaning, rather than turn every sentence into the same transaction. That requires more than fluent questions.
The useful follow-up depends on what is already known. A ride may need a destination, while tutoring may need a subject and an online or in-person preference. Asking for information that the person already gave is a failure of context. Asking endlessly is another. The engine therefore works with accepted source answers, explicit decision types and a bounded interview. The model contributes interpretation; the application keeps responsibility for how that interpretation can affect the next step.
Keep predictable decisions out of the model loop
The 17 request and offer examples now have small sets of local basic questions. Selecting the tutoring example can reveal format, timing, subjects and session length through existing controls. Online-only answers omit a place question. Offers ask about what you can do and when you are available, rather than copying the questions for someone requesting help. These basics do not make model calls until the finite local sequence ends.
This acceleration has a deliberate boundary. It requires an explicit example selection and its unchanged starting text. A handwritten request, restored draft, source edit, language change or custom note returns to the AI path so a template cannot silently reinterpret richer information. Local questions contribute confirmed answers; they do not create an authoritative assessment or declare a post ready. The final AI review and both public language versions are still required.
A model chooses from controls the platform understands
The interview has a closed set of shared controls: text, choices, number, duration, date, time, time range, location, schedule and radius. The model can request a relevant control, and the application validates the descriptor before using it. JSON Schema provides a general language for describing data structure and constraints. Our engineering lesson is that valid structure is necessary but does not establish whether a proposed question or answer is meaningful. [1]
For example, a location answer can be typed without a verified coordinate. That must remain different from an explicitly selected point on a loaded map. Likewise, “€15 per hour” cannot become “€15 for the job” just because both contain the same number. Controls and validation preserve those distinctions. Convenient suggestions remain possibilities until you choose them. A colorful tag identifies a detail in the request; it does not independently prove the real-world claim.
Corrections must change the model’s context
When you go Back and change an answer that alters the path, the engine removes the dependent answer tail and retires older AI work. The next interpretation receives the corrected source history. Private decision identities help close questions already answered or deliberately skipped. The interview stops after eight accepted turns, or earlier when no new useful decision is available. Stopping is separate from readiness: incomplete information still needs the appropriate final checks.
NIST’s AI Risk Management Framework treats trustworthiness as something to consider throughout design, development, use and evaluation. We take that as a useful discipline: separate model judgment from validation, preserve the person’s sources, and test failure paths as well as happy paths. This is our application of a general framework, not a certification or a claim that hallucination has been eliminated. [2]
Evaluate useful questions, not impressive prose
Automated tests can verify source preservation, branch cleanup, control parsing and isolated publication flows. They cannot establish that every live model question is thoughtful or that every user feels understood. Our next evaluation needs realistic requests, including ambiguous dates, social invitations, shared costs and answers that change direction. We want to examine repeated questions, unsupported assumptions, unnecessary turns and whether someone else can act on the resulting post.
A good interview should feel modest: a few decisions, room for your own words and a clear ending. Spending AI work on a question the application already knows how to ask adds little value. Spending it on an important ambiguity can be worthwhile. That is the balance we are pursuing. Try an example or write your own need, and notice whether the next question genuinely helps another person understand how to respond.
Sources and further reading
Research informs our choices. It does not prove that Neeed has achieved its goals; we still have to earn that through real use.