Guided or free-form AI

How I replaced weak prompt-first interactions with guided professional decisions and optional prompting, helping establish a product direction that later shipped and gained paying customers.

Context

The project began as an idea with the head of innovation and progressed from PoC → MVP → shipped product with paying customers.

The public example adapts the workflow to creating employee training scenarios. Users define a situation and its participants, generate alternatives and review what is suitable. The scenario protects the original domain while preserving the interaction-design reasoning.

Role: Sole Product Designer, then Design Lead with two additional designers
Evidence: Observation, interviews and client feedback
Approval: PM approved the guided model
Product status: Later shipped with paying customers
Contribution boundary: Transitioned to another project after establishing the core interaction model

Problem

The initial PoC relied heavily on free-form prompting. Observation showed that inputs were often basic, under-specified and inconsistent. Important constraints were missing, output quality was poor and a potential client was dissatisfied with the results.

The problem was concrete: users knew their work, but the interface made them responsible for translating that expertise into a complete set of instructions for the model.

Responsibility

I shaped the interaction model: how users prepared inputs, defined a scenario, reviewed choices and refined a result. I started as the sole Product Designer, then led the design direction as two additional designers joined. I later moved to another project after the core interaction model had been established; later design work included their contributions.

I recommended moving the primary interaction from describing everything in a prompt to making familiar professional decisions, while preserving optional free-form control. The PM approved the guided model.

Discovery

User interviews showed that people already had a professional vocabulary for the decisions that mattered. The interface could surface those familiar decisions instead of asking users to remember and express every constraint from scratch.

In the existing public scenario, those decisions include the audience, situation, participants and level of difficulty. Some choices, such as tone and organizational guidance, carry across exercises; others change with the scenario.

Connected source material could suggest information, but users still needed to review and approve it. Automatic preparation did not make a suggestion correct.

Delayed-delivery training scenario with objective and skills fields, a difficulty slider, and suggested audience, objective and guidance from connected source material. — Familiar choices define the training scenario, while connected source material suggests an audience, objective and key guidance. Users can edit or accept each suggestion, keeping approval separate from automatic preparation.

Decisions

1. Make familiar decisions the primary interaction

I made guidance the default. Users could select familiar terms, review a summary and then generate, with the system translating their choices into instructions.

Structured choices made missing information visible and reduced the need to reconstruct the same constraints in every prompt. The trade-off was coverage: predefined options could not anticipate every request. Optional free text preserved room for additional control.

2. Preserve approved context during refinement

A result should not require a complete restart because one part needs changing. The workflow let users keep approved decisions, adjust a section or attribute and regenerate selectively while preserving the rest of the context.

Audience, tone and organizational guidance could carry forward. Users could explore a different situation or refine one part without repeatedly explaining the shared direction.

Participant setup with a support specialist and an AI-simulated customer, beside shared audience, experience level, organizational guidance and key boundaries. — Participants are defined for this scenario, while the shared training direction stays visible alongside them. The audience, experience level and organizational boundaries carry across scenarios without needing to be entered again.

3. Make review a decision point

Generated alternatives gave users something to compare, rather than treating generation as completion. They still needed to judge whether a result suited the audience and the intended exercise.

Users could select an alternative, refine it or generate again. Reviewing individual sections supported targeted changes instead of discarding everything that already worked.

Scenario review showing the selected customer behavior, Frustrated due to travel, and alternatives including Repeat caller, alongside the sections to review before publishing. — Review happens at the level of individual scenario sections. Here, users compare customer behaviors and choose the one that fits the exercise before publishing. Generation provides alternatives; deciding what is suitable still belongs to the user.

4. Retain optional prompting

Requests for finer control showed where the structured choices reached their limits. I kept free-form input available for requirements users could not express through the guided options.

The same person could use guidance for one task and need more direct control for another. Optional prompting let the workflow support both without making prompt-writing the starting requirement.

5. Work within external API constraints

Some parameters the product needed were not directly supported by the external AI-provider APIs. I worked with the team on interaction and implementation workarounds that let users express those constraints without needing to navigate the technical limitation themselves.

Those constraints shaped what the interface could offer. My contribution was to the interaction design and the shared product decisions, rather than ownership of the engineering implementation.

6. Align the product family

I advocated for and led a design-system update to align this product with another in the same product family. The products had different login experiences, inconsistent primary-action placement and different interaction conventions. The aim was a more coherent experience across both products and stronger brand equity.

Outcome

The first, largely free-form version had left a potential client dissatisfied. We returned with a more structured and guided version. The client said the revised workflow made the process clearer and produced greater consistency with their brand, and agreed to move forward with contracting.

That sequence provided qualitative client feedback and a commercial signal. It was not a measured comparison of usability, output quality or production savings.

The concept progressed from PoC to MVP and later shipped as a product with paying customers. The later shipped product reflects contributions beyond my own.

Reflection

The project reinforced that AI interfaces should begin with the user’s professional mental model, rather than the model’s preferred input format. Prompting is valuable as an escape hatch; it does not always need to be the primary interface.

Making familiar decisions explicit gave users a clearer starting point. Preserving those decisions during refinement gave them a way to keep working without starting over.